Does the Tail Wag the Dog?

The Eect of Credit Default Swaps on Credit Risk

Marti G. Subrahmanyam Stern School of Business, New York University E-mail : msubrahm@stern.nyu.edu Dragon Yongjun Tang School of Economics and Finance, University of Hong Kong E-mail : yjtang@hku.hk Sarah Qian Wang Warwick Business School, University of Warwick E-mail : qian.wang@wbs.ac.uk December 8, 2012

We thank Viral Acharya, Edward Altman, Yakov Amihud, Sreedhar Bharath, Ekkehart Boehmer, Patrick Bolton, Dion Bongaerts, Stephen Brown, Jennifer Carpenter, Sudheer Chava, Peter DeMarzo, Mathijs van Dijk, Jin-Chuan Duan, Darrell Due, Alessandro Fontana, Andras Fulop, Iftekhar Hasan, Jingzhi Huang, Kose John, Stanley Kon, Lars-Alexander Kuehn, Anh Le, Jingyuan Li, Francis Longsta, Ron Masulis, Robert McDonald, Lars Norden, Martin Oehmke, Frank Packer, Stylianos Perrakis, Xiaoling Pu, Talis Putnins, Anthony Saunders, Lukas Schmid, Ilhyock Shim, Marakani Srikant, Rene M. Stulz, Avanidhar Subrahmanyam, Heather Tookes, Hao Wang, Neng Wang, Pengfei Ye, David Yermack, Fan Yu, Gaiyan Zhang, Xinlei Zhao, Hao Zhou, Haibin Zhu, and seminar and conference participants at CEMFI, Madrid, Cheung Kong Graduate School of Business, Beijing, City University of Hong Kong, European Central Bank, Erasmus University, Rotterdam, Hong Kong Institute for Monetary Research, Lingnan University, Hong Kong, Nanyang Technological University, National University of Singapore, NYU Stern School of Business, U.S. Oce of the Comptroller of the Currency (OCC), Ozyegin University, PRMIA (Webinar), Rouen Business School, Singapore Management University, Standard & Poors, Southwestern University of Finance and Economics, Chengdu, Tsinghua University, Beijing, University of Bristol, University of Hong Kong, University of New South Wales, University of Nottingham, Ningbo, University of the Thai Chamber of Commerce, Warwick Business School, Xiamen University, the Financial Management Association 2011 Denver meetings, the 2012 China International Conference in Finance (CICF), the 2012 European Finance Association Meetings, the 2012 Risk Management Institute conference at NUS, the 2012 UBC Winter Finance Conference, the 2012 FMA Napa Conference, the 2012 Conference of the Paul Woolley Centre for Capital Market Dysfunctionality at UTS, 2012 Multinational Finance Society Meetings, and the 2012 International Risk Management Conference, for helpful comments on previous drafts of this paper. Dragon Tang acknowledges the support and hospitality of Hong Kong Institute for Monetary Research (HKIMR) as part of the work was done when he was a HKIMR visiting research fellow.

Does the Tail Wag the Dog? The Eect of Credit Default Swaps on Credit Risk

ABSTRACT Credit default swaps (CDS) are derivative contracts that are widely used as tools for credit risk management. However, in recent years, concerns have been raised about whether CDS trading itself aects the credit risk of the reference entities. We use a unique, comprehensive sample covering CDS trading of 901 North American corporate issuers, between June 1997 and April 2009, to address this question. We nd that the probability of both a credit rating downgrade and bankruptcy increase, with large economic magnitudes, after the inception of CDS trading. This nding is robust to controlling for the endogeneity of CDS trading. Beyond the CDS introduction eect, we show that rms with relatively larger amounts of CDS contracts outstanding, and those with relatively more no restructuring contracts than other types of CDS contracts covering restructuring, are more adversely aected by CDS trading. Moreover, the number of creditors increases after CDS trading begins, exacerbating creditor coordination failure for the resolution of nancial distress.

Keywords: Credit default swaps, credit risk, bankruptcy, empty creditor

I.

Introduction

Credit default swaps (CDS) are insurance-type contracts that oer buyers protection against default by a debtor. The CDS market grew by leaps and bounds from $180 billion in 1997 to $62 trillion in 2007, measured by notional amount outstanding.1 CDS are arguably the most controversial nancial innovation of the past two decades, extolled by some and disparaged by others.2 CDS played a prominent role in the bankruptcy of Lehman Brothers, the collapse of AIG, and the sovereign debt crisis of Greece. Although the CDS market shrank considerably following the global nancial crisis, it nevertheless stood at about $29 trillion by December 2011. In spite of misgivings about the role of CDS in potentially destabilizing markets, their role as indicators of credit quality has, in fact, expanded. CDS spreads are widely quoted by practitioners and regulators for the assessment of credit risks, for both individual corporate debtors and the overall sovereign risk of a country. Meanwhile, on-shore CDS trading was launched in China and India after the credit crisis. In contrast to the intense public debate, theoretical arguments and policy initiatives, empirical evidence on the real eects of CDS trading on corporations referenced by CDS contracts is sparse. In this paper, we attempt to ll this gap in the literature, using a comprehensive dataset to empirically examine the eects of CDS on the credit risk of the reference rms. Derivatives are often assumed to be redundant securities in pricing and hedging models and hence have no eect, adverse or benign, on the price of the underlying asset or the integrity of markets. In structural models of credit risk along the lines of Merton (1974), default risk is driven principally by leverage and asset volatility. In the spirit of that framework, CDS are regarded as side-bets on the value of the rm and hence do not have an impact on the credit risk associated with the individual claims issued by the rm. In particular, in such models, CDS trading does not aect the probability of bankruptcy or even the possibility of a credit rating downgrade. Many of the issues mentioned in the context of derivatives, in general, have also been raised
Semiannual OTC Derivative Statistics, Bank for International Settlements (BIS). CDS market statistics are also regularly published by the International Swaps and Derivatives Association (ISDA) and the British Bankers Association (BBA). 2 Former Federal Reserve Chairman Alan Greenspan argued that these increasingly complex nancial instruments have contributed, especially over the recent stressful period, to the development of a far more exible, ecient, and hence resilient nancial system than existed just a quarter-century ago. (See Economic Flexibility, Alan Greespan, Speech given to Her Majestys Treasury Enterprise Conference, London, January 26, 2004.) In striking contrast, Warren Buett, the much-acclaimed investor, weighed against derivatives, in general, by describing them as time bombs, for the parties that deal in them and the economic system and went on to conclude that in my view, derivatives are nancial weapons of mass destruction, carrying dangers that, while now latent, are potentially lethal. (See the Berkshire Hathaway Annual Report for 2002.) In a similar vein, George Soros, a legendary hedge fund manager, argued that CDS are toxic instruments whose use ought to be strictly regulated. (See One Way to Stop Bear Raids, Wall Street Journal, March 24, 2009.)
1

in the specic case of CDS regarding their eect on the underlying asset.3 Apart from common concerns that apply to all derivatives, CDS contracts are somewhat dierent. CDS contracts are traded over-the-counter, where price transparency and discovery are less clear-cut than in the exchanges on which most equity derivatives are listed. Moreover, nancial institutions, including the bank creditors of the reference entities, are major participants of the CDS market. CDS typically have much longer maturities than most exchange-traded derivatives, allowing the traders more exibility in adjusting their positions. If creditors selectively trade CDS linked to their borrowers, CDS positions can change the creditor-borrower relationship and play an important role in determining the borrower credit risk that determines CDS payos. On the one hand, CDS allow creditors to hedge their credit risk; therefore they may increase the supply of credit to the underlying rm. Such improved access to capital may increase borrowers nancial exibility and resilience to nancial distress.4 On the other hand, lenders may not be as vigilant in monitoring the borrowers once their credit exposures are hedged. Consequently, rms, in turn, may take on more risky projects. Furthermore, CDS-protected creditors are likely tougher during debt renegotiations, once the borrowers are in nancial distress, by refusing debt workouts and making borrowers more vulnerable to bankruptcy. We empirically examine the eects of CDS trading on the credit risk of reference entities using a comprehensive dataset dating back to the broad inception of the CDS market for corporate names in 1997. It should be emphasized that it is dicult to obtain accurate data on CDS transactions from a single source, since CDS trading does not take place on centralized exchanges. Indeed, the central clearing of CDS is a relatively recent phenomenon. Our identication of CDS inception and transactions relies, of necessity, on multiple data sources including GFI Inc., the largest global interdealer broker with the most extensive records of CDS trades and quotes, CreditTrade, a major intermediary especially in the early stages of the CDS market, and Markit, a data disseminator and vendor that provides daily valuations based on quotes from major sell-side institutions. Our combined dataset covers 901 North American rms with a CDS trading history during the period from 1997 to 2009. The list of bankruptcies for North American rms is comprehensively constructed from major data sources such as New Generation Research, the UCLA-LoPucki Bankruptcy Database, the Altman-NYU Salomon Center Bankruptcy List, the Fixed Income Securities Database (FISD), and Moodys Annual Reports on Bankruptcy and Recovery. Over the same time
At a general level, there is evidence from the equity market that derivatives trading can aect the pricing of the underlying asset. See, for example, an early survey by Damodaran and Subrahmanyam (1992), and Sorescu (2000), for examples of such studies. 4 Indeed, this argument has been cited as the motivation for the invention of CDS by JPMorgan, which lent to Exxon Mobil in 1994 in the aftermath of the Exxon Valdez oil spill lawsuit. In a pioneering transaction, JPMorgan hedged part of its credit exposure using a CDS transaction with the European Bank for Reconstruction and Development (EBRD). See Tett (2009).
3

period, our overall sample of rms covers 3,863 rating downgrades from Standard & Poors and 1,628 bankruptcy lings. Our rst nding from the combined dataset is that, controlling for fundamental credit risk determinants suggested by structural models, the likelihood of a rating downgrade and the likelihood of the bankruptcy of the reference rms both increase after CDS start trading. The increase in credit risk after CDS trading begins is both statistically signicant and economically meaningful. For our sample of CDS rms, credit ratings decline by about half a notch, on average, in the two years after the inception of CDS trading. In a similar vein, the probability of bankruptcy more than doubles (from 0.14% to 0.47%) once the CDS start trading on a rm. The selection of rms for CDS trading and the endogeneity of the timing of CDS inception need to be addressed in order to make a causal inference about the eect of CDS trading. CDS rms and non-CDS rms are quite dierent in terms of their key characteristics. There could be unobserved omitted variables that drive both the selection of rms for CDS trading and changes in bankruptcy risk. Also, the timing of CDS inception can be endogenous as CDS trading is more likely to be initiated when market participants anticipate the future deterioration in the credit quality of the reference rm. We address these two concerns in several ways besides the basic xed eects controls. Specically, we construct a model to predict CDS trading for individual rms. This model allows us to measure the treatment eect of CDS inception using an instrumental variable (IV) approach, run a propensity score matching analysis for rms with and without CDS trading, and conduct a dierence-indierence estimation. We nd two IVs for CDS trading. The rst IV is the foreign exchange (FX) hedging position of lenders and bond underwriters. Lenders with a larger FX hedging position are more likely, in general, to trade the CDS of their borrowers. The second is the lenders Tier One capital ratio. Banks with lower capital ratios have a greater need to hedge the credit risk of their borrowers via CDS. It seems valid to exclude both IVs from the credit risk predictions of rms since they only aect borrower credit risk via CDS market activities. We also show that both IVs are signicant determinants of CDS trading and that they are not weak instruments. Furthermore, the Sargan over-identication tests fail to reject the hypothesis that both IVs are exogenous. The positive relationship between CDS trading and bankruptcy risk remains signicant, even after controlling for the selection and endogeneity of CDS trading. The eect of CDS trading on credit risk goes beyond the simple binary categorization of rms CDS status. It is conceivable that CDS will be more inuential when the market is more liquid and when more contracts are outstanding. Indeed, we nd that the likelihood of bankruptcy increases with the number of live CDS contracts outstanding. Therefore, the eect of CDS works in both directions: Bankruptcy risk increases as CDS positions gather force 3

and decreases when the amount of CDS trading is reduced. These ndings further strengthen the evidence that the increase in credit risk after CDS trading begins is not completely due to selection and endogeneity. After establishing our primary nding that the reference rms credit risk increases after CDS trading begins, we investigate potential mechanisms for channeling the eect of CDS trading on credit risk. CDS can aect rm fundamentals such as the leverage and the interest burden. The credit risk of a rm clearly increases as it becomes more leveraged. Indeed, we nd that rm leverage increases signicantly after CDS trading begins. The increase in leverage can be due to either enlarged credit supply or reduced debt nancing restrictions imposed by lenders after CDS trading has begun.5 Therefore, we control for leverage (both before and after CDS trading) in our regression analysis in order to isolate the leverage channel from other possibilities. The credit risk of a rm can also increase if it is more vulnerable in nancial distress. One source of vulnerability arises from the creditors unwillingness to work out troubled debt. Another source is the potential failure of coordination among the distressed rms creditors. Anecdotal evidence suggests that CDS positions can play an important role in the process of distress resolution. To cite one such instance, CIT Group attempted to work out its debt from late 2008 to mid-2009. In the event, however, some creditors with CDS protection rejected the rms exchange oer.6 CIT Group eventually led for Chapter 11 bankruptcy on November 1, 2009. Hu and Black (2008) term such CDS-protected debt-holders empty creditors, meaning that they have all the same legal rights as creditors, but do not have positive risk exposure to borrower default; hence, their nancial interests are not aligned with those of other creditors who do not enjoy such protection.7 The empty creditor problem is formally modeled by Bolton and Oehmke (2011).8 Their model predicts that, under mild assumptions, lenders will choose to become empty creditors by buying CDS protection. Consequently, they will be tougher in debt renegotiation when the rm is under stress. Empty creditors are even willing to push the rm into bankruptcy if their total payos including CDS payments would be larger in that event. In their model, CDS sellers anticipate this empty creditor problem and price it into the CDS premium, but
Saretto and Tookes (2012) focus on the eect of CDS trading on leverage and conrm the hypothesis of increased leverage. 6 See Goldman Purchase Puts CDS in Focus, Financial Times, October 4, 2009, and Goldman Sachs May Reap $1 Billion in CIT Bankruptcy, Bloomberg, October 5, 2009. 7 The use of equity derivatives such as options or swaps in the context of equities creates the analogous issue of empty voters who enjoy voting rights in the rm, but without any nancial risk, by breaking the link between cash ow rights and control rights. 8 Table 1 of Bolton and Oehmke (2011) lists other cases of suspected empty creditors, demonstrating that the CIT example is not that unique. Other studies such as Due (2007), Stulz (2010), and Jarrow (2011) also oer relevant discussions on creditor incentives.
5

they cannot directly intervene in the debt renegotiation process (unless they buy bonds or loans so as to become creditors). Our data do not include trader identities; therefore, we cannot directly observe the presence and extent of empty creditors; neither are we aware of other data sources that would allow direct detection of empty creditors. In an indirect test, we nd that rm bankruptcy risk is positively related to the total CDS amount divided by total debt. We further construct a more eective test of tough creditor implications. Our combined dataset contains contract terms that allow us to test a unique prediction of the empty creditor model. Specically, we know for each CDS contract whether restructuring is covered as a credit event or not. Buyers of no restructuring CDS contracts will be paid only if the reference rm les for bankruptcy or there is a failure to pay. However, buyers of other types of CDS contracts that include restructuring as a credit event will be compensated even when the debt of the reference rm is restructured. Clearly, creditors with no restructuring CDS protection will have a stronger incentive to force bankruptcy than buyers of other CDS contracts without this restrictive clause. Indeed, we nd that the eects of CDS trading are stronger when a larger fraction of the CDS contracts contain the no restructuring credit event clause. This result also provides evidence of the causal eects of CDS trading, particularly since there is no signicant eect from other types of CDS contracts and, even more so, since this measure does not directly rely on the selection of rms for CDS trading. The availability of CDS contracts may render more banks willing to lend, due to the possibility of risk mitigation and enhanced bargaining power via CDS contracts. However, such an expanded lender base can also hinder debt workouts. The greater the number of lenders, the more likely that some lenders will choose to become empty creditors, and the more severe will be the problems of coordination in a stressed situation, when a workout may be necessary. Therefore, CDS trading may aect lending relationships, and in particular the number of lenders. Indeed, we nd that more creditors lend to the rms after reference CDS become available. Consistent with prior ndings, we also nd that bankruptcy risk increases with the number of lenders due to creditor coordination failure, thus providing another channel for the adverse eect of CDS trading on bankruptcy risk. In sum, rather than being an instrument providing insurance against borrower default, CDS trading can increase the likelihood of borrower default (the tail wags the dog). Our main contribution is documenting a real eect of the trading of CDS on the survival probabilities of rms. We are among the rst to formally test and support the empty creditor model of Bolton and Oehmke (2011). Our study complements Ashcraft and Santos (2009) and Saretto and Tookes (2012), who nd that the cost of debt of risky rms, and their leverage, increase after CDS trading has started.

Our ndings have implications for investors in credit markets as well as rms. These entities need to consider the impact of CDS trading on the likelihood of bankruptcy in their pricing of corporate debt. Financial regulators and policy makers need to take the increase in credit risk following CDS trading into account in their regulatory actions. In particular, banking regulators need to incorporate this eect in their risk weighting formulae, while securities regulators may require further disclosures of CDS positions, so that investors are made aware of the extent of the potential impact of CDS trading on credit risk. The remainder of this paper is organized as follows. Section II develops testable hypotheses in relation to the literature. The construction of our dataset is described in Section III. Section IV presents our empirical results for the eect of CDS trading along with a detailed examination of the endogeneity concerns and the mechanisms for the eect. Section V concludes.

II.

Related Literature and Testable Hypotheses

CDS were originally invented to help banks to transfer credit risk, maintain relationships with borrowers, and expand their business. The availability of CDS has indeed aorded banks the exibility and opportunity to manage their credit risk. Over time, other participants such as hedge funds and mutual funds have become active in the CDS market. We place our research in the context of the literature on the CDS market, with particular reference to studies that address issues relating to the relationship between rms and their creditors.9 Several recent theoretical studies model the role of CDS in debt nancing. Bolton and Oehmke (2011) argue that credit supply can increase because creditors will be tougher and have more bargaining power in debt renegotiation when they use CDS to protect their exposure, thereby reducing borrowers incentives for strategic default. On the other hand, Che and Sethi (2012) conjecture that CDS can crowd out lending as creditors can sell CDS instead of making loans or buying bonds, eectively reducing credit supply and increasing the cost of debt. Campello and Matta (2012) point out that the eect of CDS depends on macroeconomic conditions. The empirical evidence relating to the eect of CDS on the cost and supply of debt is mixed. Ashcraft and Santos (2009) nd that, after CDS introduction, the cost of debt increases for low-quality rms and decreases for high-quality rms. While Hirtle (2009) nds no signicant increase in bank credit supply after the initiation of CDS trading, Saretto
There is a vast literature on other aspects of CDS trading. Longsta, Mithal, and Neis (2005), Stanton and Wallace (2011), and Nashikkar, Subrahmanyam and Mahanti (2012) discuss the pricing of CDS. Apart from individual rms in the economy, CDS trading may also have an eect on the aggregate economy. For instance, Duee and Zhou (2001) and Allen and Carletti (2006) show that CDS trading may hurt nancial stability when rms are interconnected. Arping (2004) and Morrison (2005) argue that CDS can reduce the lender-borrower combined welfare.
9

and Tookes (2012) nd that the reference rms leverage increases. There are potentially both positive and negative inuences of CDS trading on the credit risk of reference entities. On the one hand, if the leverage of a rm increases after CDS trading has begun, it follows that its bankruptcy risk also increases correspondingly. Moreover, as we illustrate in Appendix A, the lenders willingness to restructure the rms debt in the event of nancial distress is aected by their respective CDS positions. Some CDS-protected lenders may prefer the bankruptcy of borrowers, if the payos from their CDS positions are high enough. Although there are other reasons why lenders may be unwilling to restructure the debt of a rm in nancial distress (for example, they may believe that the borrower could eventually go bankrupt even after a debt restructuring), their CDS positions will be a factor in their decision. On the other hand, issuers could benet from CDS trading on their names. Allen and Carletti (2006) show that, under certain conditions, CDS improve risk sharing and are good for both borrowers and lenders. Parlour and Winton (2012) construct a model showing that CDS can help improve lending eciency for high-quality borrowers. Norden, Silva-Buston, and Wagner (2012) show that lenders with more CDS activities oer lower loan rates and help their borrowers during periods of nancial crisis. It follows that, if CDS are benecial to the lenders, then some of the benets may be passed on to or shared with the borrowers, thus making rms safer. If the risks outweigh the benets of nancial exibility, then we expect rms to be riskier after CDS trading: Hypothesis 1 (Baseline) The credit risk of a rm and, in particular, its risk of bankruptcy increase after the introduction of trading on CDS contracts referencing its default. One could alternatively examine the related hypothesis that CDS trading reduces the success rate of restructuring for distressed rms. This latter question has been addressed in three complementary studies, albeit with smaller samples, by Bedendo, Cathcart, and El-Jahel (2012), Danis (2012) and Narayanan and Uzmanoglu (2012), with conicting conclusions. While Danis (2012) nds signicant impact of CDS trading on restructuring, Bedendo, Cathcart, and El-Jahel (2012) and Narayanan and Uzmanoglu (2012) fail to nd such eects. Our analysis applies to the full sample of rms, both healthy and distressed. Bankruptcy may be a better testing framework than restructuring as bankruptcy events are more easily observed than restructuring events. Moreover, dening distressed rms in the context of restructuring is a subjective assessment, which poses challenges for the researcher (and may explain the mixed evidence from above-mentioned studies). Therefore, we focus on bankruptcy lings in our analysis here. The eect of CDS trading can vary considerably even among CDS rms. Indeed, Minton, Stulz, and Williamson (2009) nd that banks use of CDS depends on the market liquidity of the particular instrument. The larger is the holding of CDS relative to debt outstanding, 7

the greater is the benet to CDS buyers, and hence, their incentive to tilt the rm towards bankruptcy. Therefore, we quantify the CDS eect based on the amount of CDS trading in the following hypothesis: Hypothesis 2 (CDS Exposure) The increase in the bankruptcy risk of a rm after the introduction of trading in CDS contracts referencing its default is larger for a rm with a greater number of CDS contracts outstanding. Another distinctive feature of our study is that we test for the quantitative implications of CDS trading. Peristiani and Savino (2011) document that higher bankruptcy risk is signicant in the presence of CDS during 2008, but insignicant overall in their sample. Our study uses a comprehensive database and rigorous econometric procedures to provide more powerful tests than the binary CDS introduction events. We next address the issue of the mechanisms by which CDS trading aects bankruptcy risk, with particular emphasis on the incentives of tough creditors.10 Empty creditors do not completely determine the fate of the reference entities. In some cases, the reference rms survive without any credit events, or with straightforward debt rollover, if other creditors support the borrower and outweigh the inuence of empty creditors. In such cases, empty creditors will lose the additional premium they paid to the CDS sellers without any concomitant benets. However, if credit events do occur, empty creditors and other CDS buyers will likely make prots. (Thompson (2010) shows that the insurance buyer will also need to worry about whether the seller can honor its commitment.) Whether the overall eect of CDS trading is signicant or not depends on the incentives of the marginal creditors, and will be borne out in the data. If we can make the assumption that the presence of CDS implies a higher probability of empty creditors than there are for non-CDS rms, then our primary hypothesis will also answer this question. Moreover, we take advantage of information on the amount of CDS relative to debt outstanding and the presence of the restructuring clause in the CDS contracts: Hypothesis 3 (Tough Creditors) The increase in the bankruptcy risk of a rm after the introduction of trading in CDS contracts on it is larger if (a) there is a greater notional amount of CDS contracts relative to debt outstanding (over-insurance), and (b) no restructuring (NR) contracts account for a larger proportion of all CDS contracts referencing its default. The third hypothesis suggests a unique test of the empty creditor mechanism by using a special feature of the CDS contracts. If CDS contracts cover restructuring as a credit event, then
One natural related question is: Are creditors tougher under CDS trading? The recent decline in the absolute priority deviation during bankruptcy resolution documented by Bharath, Panchapagesan, and Werner (2010) is consistent with tougher creditors and coincides with the development of the CDS market. However, this issue merits more detailed investigation.
10

creditors will be compensated, whether the distressed rm restructures or declares bankruptcy. However, if restructuring is not covered in the restructuring clause, the default event may be triggered, but the empty creditor will only get compensated if there is a failure to pay or the rm les for bankruptcy. Therefore, we hypothesize that the empty creditor mechanism is even more eective for NR CDS. We note that Bolton and Oehmke (2011) endogenize the pricing of CDS contracts so that the CDS seller takes this empty creditor incentive into account. The hypothesis above emphasizes the ex post eect (after the loan and CDS positions are given) of CDS due to lenders that are tougher in debt renegotiation, although not every creditor would want to become an empty creditor. Gopalan, Nanda, and Yerramilli (2011) show that the lead bank suers reputation damage from borrower bankruptcies. From an ex ante perspective, lenders could be strategic in their use of CDS and lending decisions. Bolton and Oehmke (2011) show that lenders are more willing to lend when CDS permit them the possibility of risk mitigation. It follows that more banks are willing to lend to a rm when CDS are available.11 Such an expansion in the lender base and the level of lending has two consequences. First, the likelihood of empty creditors is higher when there are more lenders. Second, the probability of bankruptcy is higher when there are more lenders due to the potential for coordination failure. Gilson, John, and Lang (1990) show that creditor coordination failure increases the risk of bankruptcy. Brunner and Krahnen (2008) show that distress workouts are less successful when there are more creditors. Therefore, we generate our last hypothesis in two parts: Hypothesis 4 (Lender Coordination) (a) The number of (bank) lenders increases after the introduction of CDS trading. (b) Bankruptcy risk increases with the number of lenders.

III.

Dataset on CDS Trading and Bankruptcy

We use actual transaction records to identify rms with CDS contracts written on them, and in particular, the date when CDS trading began for each rm and the type of contract traded. Unlike voluntary dealer quotes that are non-binding and may be based on hypothetical contract specications, transaction data contain multi-dimensional information on the actual CDS contracts, including price, volume and settlement terms. Our CDS transactions data are obtained from two separate sources: CreditTrade and GFI Group. CreditTrade was the
Borrowers may also want to broaden their lender base if they anticipate that some lenders could take advantage of their respective CDS positions. Acharya and Johnson (2007) suggest that bank lenders engage in insider trading in the CDS market. Hale and Santos (2009) show that, if banks exploit their information advantage, rms respond by expanding their borrowing base to include lenders in the public bond market or by adding more bank lenders.
11

main data source for CDS transactions during the initial phase of the CDS market, before GFI Group took over as the market leader.12 Combining data from these two sources allows us to assemble a comprehensive history of North American corporate CDS trading activities. Our CreditTrade data cover the period from June 1997 to March 2006, while our GFI data cover the period from January 2002 to April 2009. Both datasets contain complete information on intra-day quotes and trades such as the type of contract, the time of the transaction, order type, and the CDS price. Since CDS contracts are traded over-the-counter, unlike stocks or equity options, which are mostly traded on exchanges, the rst trading date for each rms CDS is hard to pinpoint with a time stamp. However, because we have overlapping samples from these two data sources between January 2002 and March 2006, we are able to cross-check the two records to conrm the reliability of our identication of the rst CDS trading date. In the event, the dates of rst appearance of a particular CDS in the two data sources are mostly within a couple of months of each other. To ensure greater accuracy, we also crosscheck trading-based CDS data with the Markit CDS database, a commonly used CDS dealer quote database, and conrm our identication of rms for which CDS are traded and the date of inception of trading.13 It should be stressed that any remaining noise in identifying the precise introduction date of a particular CDS should bias us against nding signicant empirical results regarding the consequent eects on credit risk. There are two important advantages of using the complete set of transaction data in our empirical analysis of non-sovereign North American corporate CDS. First, our sample starts in 1997, which is generally acknowledged to be the year of inception of the broad CDS market.14 Therefore, our identied rst CDS trading dates will not be contaminated by censoring of the data series. Second, our CDS transaction data include the complete contractual terms, such as the specication of the credit event, maturity, and security terms, at the contract level. Aggregate position or quote data obtained from broker-dealers or, more recently, clearing houses or data aggregators, would generally not include such detailed information. The credit event specication allows us to investigate the eect of restructuring clauses. The maturity information at the contract level allows us to calculate the amount of the outstanding CDS positions at each point in time. Our sample of CDS introductions ends in April 2009 for an important institutional reason: The market practice in CDS changed signicantly in April 2009 due to the Big Bang implemented by ISDA, including for example the removal of restructuring as a standard credit event. In addition, we need an observation window of three
12 Previous studies have used the same data sources. For example, Acharya and Johnson (2007) and Blanco, Brennan, and Marsh (2005) utilize CreditTrade data. Nashikkar, Subrahmanyam, and Mahanti (2011) use CDS data from GFI. GFI ranked rst in the Risk Magazine CDS broker ranking from 2006-2010. (CreditTrade was acquired in 2007 by Creditex, which merged with the CME in 2008.) 13 Markit provides end-of-day average indicative quotes from contributing sell-side dealers, using a proprietary algorithm. In contrast, both CreditTrade and GFI report trades as well as binding quotes. 14 See Tett (2009) for a historical account.

10

years after the introduction of CDS trading to capture its potential eects in our empirical analysis. Based on our merged dataset, there are 901 North American rms that have CDS initiated on them at some point during the 1997-2009 sample period. The industry distribution of the CDS rms in our sample is quite diverse.15 In our baseline analysis, we mainly utilize the information about the rst day of CDS trading, and compare the changes in rm default risk upon the onset of CDS trading. Later on, we also construct measures of the amount of CDS outstanding and the fraction of CDS contracts with various restructuring clauses, based on more detailed transaction information, to further understand how CDS trading aects credit risk. We assemble a comprehensive bankruptcy dataset by combining data from various sources for North American corporations ling bankruptcies in U.S. courts. Our initial bankruptcy sample is derived from New Generation Researchs Public and Major Company Database, available at www.BankruptcyData.com. This database includes information on public companies ling for bankruptcy and also signicant bankruptcies of private rms. We further validate and augment this initial sample with additional bankruptcy-ling data sources, including the Altman-NYU Salomon Center Bankruptcy List, the Mergent Fixed Income Securities Database (FISD), the UCLA-LoPucki Bankruptcy Research Database (BRD), and Moodys Annual Reports on Bankruptcy and Recovery. We use Dealscan Loan Pricing Corporation (LPC) and FISD data to identify the lenders and underwriters to a rm. We obtain data on foreign exchange hedging from the Federal Reserve call reports and bank capital ratio data from the Compustat Bank le. Our rm data are drawn from the Compustat database. Our sample covers bankruptcies of both large and small rms (many studies in the literature only examine large rms). We link the bankruptcy dataset with our CDS sample to identify the bankrupt rms that had CDS trading prior to their bankruptcy lings. Table I presents the yearly summary from 1997 to 2009 for all rms in the Compustat database: the number of bankrupt rms, the number of rms on which CDS are traded, and the number of bankrupt rms with and without CDS trading. The last row of Table I shows a total gure of 1,628 bankruptcy lings during the 1997-2009 sample period. Many bankruptcies were led in the periods of 1999-2003 and 2008-2009, accounting for 1,214 of the 1,628 bankruptcy events during the entire sample period (74.6%). The fourth and fth columns of the table report the number of New CDS rms and the number of rms with Active CDS trading rms across the years, respectively. More CDS contracts were introduced in the period 2000-2003 than in earlier or
Most CDS rms in our sample are in the manufacturing (SIC 2, 3), transportation, communications, and utilities (SIC 4), and nance, insurance, and real estate (SIC 6) sectors. In our empirical analysis, we control for industry xed eects throughout.
15

11

later periods. Among the 901 distinct CDS trading rms, 60 (6.7%) subsequently led for bankruptcy protection. Bankruptcies among CDS rms represent a small fraction of the total number of bankruptcies, since only relatively large rms, by asset size and debt outstanding, have CDS trading. However, the bankruptcy rate of 6.7% for CDS rms is close to the 4year overall (or 11-year BBB-rated) cumulative default rate of U.S. rms (Standard & Poors (2012)).

IV.

CDS Trading and Credit Risk: Empirical Results

This section presents our empirical ndings on the eect of CDS trading on a rms credit risk. We use several common measures of credit risk, including credit rating, probability of bankruptcy, and expected default frequency, in our analysis. First, we report our baseline results on the eects of the introduction of CDS trading. Second, we address the issue of selection and endogeneity in the introduction of CDS trading. Third, we examine the eect of CDS positions and contract terms, and investigate the mechanisms through which CDS trading aects credit risk.

A. Rating Distributions Before and After CDS Introduction

A straightforward ordinal measure of credit risk is the credit rating that is widely used in industry. We study the characteristics of CDS rms by rst analyzing their credit ratings around the time of the introduction of CDS trading. If the issuer credit quality changes after the introduction of CDS trading, the credit ratings may reect this CDS eect if rating agencies perform reasonable credit analysis. Rating agencies incorporate information on both bankruptcies and restructuring into rating decisions (Moodys (2009)). In addition, since a credit rating downgrade is often the rst step towards bankruptcy and is an indicator of an increase in bankruptcy risk, it may convey useful information about the probability of bankruptcy. We obtain the time series of Standard & Poors (S&P) long-term issuer ratings from Compustat and FISD. We then conduct an event study of the eect of the introduction of CDS trading on credit ratings to gain a high-level understanding of the evidence. This is a basic within-rm analysis, in which we compare the distribution of credit ratings in the year right before CDS trading (year t 1), with the rating distribution two years after CDS trading has begun (year t + 2), for all rms with such contracts traded at some point in our sample. These rating distributions, one year before and two years after the introduction of CDS trading, are plotted in Figure 1. Our rst observation from Figure 1 is that A and BBB ratings are the most common issuer ratings at the time when CDS trading is initiated. 12

The vast majority of rms in our sample (92%) are rated by a credit rating agency at the onset of CDS trading, with only a small proportion of rms being unrated at this juncture. Compared to the general corporate rating distribution documented in Grin and Tang (2012), our sample includes more BBB-rated rms relative to other investment grade (AAA, AA, Arated) rms, but also has fewer non-investment grade rms. Overall, rms in our sample are of relatively good credit quality, as measured by credit ratings, at the time of CDS inception. Figure 1 shows a discernible shift to lower credit quality after the introduction of CDS trading. While the proportion of BBB-rated rms is about the same before and after CDS trading begins, the proportion of AA-rated and A-rated rms decreases. At the same time, the proportion of non-investment grade and unrated rms increases. The Kolmogorov-Smirnov test statistic for the distributional dierence before and after CDS trading begins is signicant at the 1% level, indicating that the credit rating distribution shifts to the right (lower rating quality) after CDS trading begins. Specically, 54% of the rms maintain the same ratings before and after the introduction of CDS trading, 37% of the rms experience rating downgrading but only 9% of rms experience a rating improvement.16 These results provide preliminary evidence that the credit quality of the reference entities deteriorates following the inception of CDS trading.

B. Baseline Hazard Model Results on Downgrading and Bankruptcy

We next run multivariate tests to discern systematic statistical evidence, with appropriate control variables, regarding the eect of the inception of CDS trading on credit risk. We include rms with and without CDS traded in a panel data analysis, using monthly observations. We examine both credit rating downgrades and bankruptcy lings in our baseline analysis. There is a large literature on bankruptcy prediction dating back to the Z -score model of Altman (1968). Bharath and Shumway (2008) and Campbell, Hilscher, and Szilagyi (2008) discuss the merits of simple bankruptcy prediction models over their more complicated counterparts and argue that the simple models perform quite well in predicting bankruptcy. In keeping with this perspective, our approach is a proportional hazard model for bankruptcy using panel data.17 Following Shumway (2001), Chava and Jarrow (2004), and Bharath and Shumway (2008), we assume that the marginal probability of bankruptcy over the next period
16 We also nd that, compared to non-CDS rms from the same industry and of similar size, there are 2.6% more rating downgrades for CDS rms after CDS trading starts than for non-CDS rms at the same time. 17 We also perform robustness checks on this model specication later on.

13

follows a logistic distribution with parameters (, ) and time-varying covariates Xit1 : Pr(Yit = 1|Xit1 ) = 1 , 1 + exp( Xit1 ) (1)

where Yit is an indicator variable that equals one if rm i les for bankruptcy in period t, and Xit1 is a vector of explanatory variables observed at the end of the previous period. A higher level of + Xit1 represents a higher probability of bankruptcy. We follow Bharath and Shumway (2008) to include ve fundamental determinants of default risk in Xit1 : the logarithm of the rms equity value (ln(E)), the rms stock return in excess of market returns over the past year (rit1 rmt1 ), the logarithm of the book value of the rms debt (ln(F)), the inverse of the rms equity volatility (1/E ), and the rms protability measured by the ratio of net income to total assets (NI/TA).18 We obtain rm accounting and nancial data from CRSP and Compustat. In addition to these ve fundamental variables we include two CDS variables, CDS Firm and CDS Active, in the hazard model specications to estimate the impact of CDS trading on bankruptcy risk, similarly to Ashcraft and Santos (2009) and Saretto and Tookes (2012). CDS Firm is a dummy variable that equals one for rms with CDS traded at any point during our sample period. It is a rm xed characteristic and does not change over time. CDS Firm is used to control for unobservable dierences between rms with and without CDS. CDS Active is a dummy variable that equals one after the inception of the rms CDS trading and zero before CDS trading. CDS Active equals zero for non-CDS rms. Hence, the coecient of interest is that of CDS Active, which captures the marginal impact of CDS introduction on bankruptcy risk. Since the variables CDS Firm and CDS Active are positively correlated, we report results both with and without the control of CDS Firm in our main analysis. We also control for year and industry xed eects in the panel data analysis. We apply the same specication to the analysis of the probability of a rating downgrade. The proportional hazard model estimation results are presented in Table II. We follow Shumway (2001) and correct the standard errors by the average number of observations per cross-sectional unit. We report heteroscedasticity-consistent standard errors throughout, as in Bharath and Shumway (2008). In the proportional hazard model, the cross-sectional dependence is separated from the time-series dependence. However, this is essentially a crosssectional estimation, as the form of the baseline hazard on each date is not specied. Therefore, a heteroscedasticity correction should suce for the calculation of standard errors. The rst column lists the independent variables in the model estimation. The dependent variable for Specications 1 and 2 is the probability of a credit rating downgrade in the observation
18 Longsta, Giesecke, Schaefer, and Strebulaev (2011) argue that factors suggested by structural models, such as volatility and leverage, predict bankruptcy better than other rm variables.

14

month. The dependent variable for Specications 3 and 4 is the probability of a bankruptcy ling in the observation month. The coecient estimate for CDS Active is positive and signicant for all four specications. The eect of CDS Active is not driven by fundamental dierences between CDS rms and non-CDS rms. Specications 2 and 4 show that the eect of CDS Active is signicant, even without controlling for CDS Firm. The coecient estimates for the variable CDS Firm are statistically signicant at the 1% level in both Specication 1 and Specication 3, but with opposite signs. That is, compared to non-CDS rms, CDS rms are, in general, more likely to be downgraded but less likely to go bankrupt. Such a diametrically opposite eect of CDS Firm is in contrast to the consistently positive CDS Active eect, further attenuating the concern that the eect of CDS Active is driven by multi-collinearity with CDS Firm. The positive coecients of CDS Active in Specications 1 and 2 indicate that rms are more likely to be downgraded after the inception of CDS trading. In both specications, the eect of CDS trading is statistically signicant at the 1% level. The economic magnitude is also large: Compared to the average downgrading probability of 0.58% in Specication 1, the marginal eect of CDS trading on the probability of a downgrade is 0.39%. Specication 3 reports similar ndings for bankruptcy ling. Bankruptcy risk increases after CDS trading has begun: Against an average rm bankruptcy probability of 0.14%, the marginal eect of CDS trading on the bankruptcy probability is 0.33%. The odds ratio for CDS Active (the likelihood of downgrading/bankruptcy after CDS trading divided by the likelihood of downgrading/bankruptcy before CDS trading) for credit downgrades and bankruptcy predictions are 1.925 and 10.73 respectively, indicating that credit events are much more likely after CDS trading begins. The eect of CDS Active is not driven by industry eects as we control for them throughout our analysis. The estimation results for the other control variables in Table II are similar to the ndings in prior studies. Larger rms and rms with higher stock returns are less likely to be downgraded or to go bankrupt. Firms with higher leverage and greater equity volatility are more likely to be downgraded or go bankrupt, all else being the same. As is to be expected, protable rms are less likely to le for bankruptcy. Lastly, the pseudo-R2 s, about 15% for the downgrade regressions and 24% for the bankruptcy regressions, suggest that bankruptcy lings are better explained by these explanatory variables than downgrades. In sum, Table II of our baseline analysis shows consistent results that the credit quality of reference rms declines after CDS trading begins. We also run a battery of robustness checks on our baseline results for bankruptcy ling. First, we consider rm xed eects rather than industry xed eects. We cannot include rm xed eects in our bankruptcy analysis as the estimation does not converge due to its nonlinear specication. Therefore, we use distance-todefault as the dependent variable. Such a specication allows us to include rm xed eects 15

(in this case we do not need to include the CDS Firm control). Moreover, we show that our ndings are robust to alternative model specications, rating drift consideration and other rm exits.19 Next we present the results of several alternative approaches, used to address the selection and endogeneity concerns in CDS trading.

C. Selection and Endogeneity in CDS Trading

The previous subsection shows a strong relation between CDS trading and the subsequent increase in credit risk. However, the main challenges to inferring a causal relationship showing that CDS trading leads to a deterioration in credit quality are the potential selection and endogeneity in CDS trading. Selection eects would be a concern if CDS rms were fundamentally dierent from non-CDS rms, and such fundamental dierences were related to the subsequent deterioration in credit quality. Nevertheless, the selection of rms into the CDS sample may not be our biggest concern, since our focus is on the timing of CDS trading.20 Essentially, we are interested in the within-rm eect, where the timing of the introduction of CDS trading may be endogenous. It is conceivable that CDS traders anticipate the deterioration in a rms credit quality and initiate trading in its CDS contract. Therefore, CDS Active, the variable measuring the eect of the timing of CDS introduction, is the main endogenous variable of concern. We note that examining CDS Active for endogeneity also takes CDS Firm (the selection of rms into the CDS sample) into account as CDS Active is always zero for non-CDS rms. We use several standard econometric approaches to address the endogeneity and selection issues, as suggested by Li and Prabhala (2007) and Roberts and Whited (2012): IV estimation, the Heckman treatment eects model, propensity score matching, and dierence-in-dierence estimation. We need to rst have a good understanding of the determinants of CDS trading before we can eectively apply the various econometric approaches to address the endogeneity and selection issues. We aim to nd the most appropriate model for the selection of CDS trading on rms, so that we can then adjust for this selectivity in our analysis of credit risk changes after the start of CDS trading. We follow Ashcraft and Santos (2009), Saretto and Tookes (2012), and other studies with similar endogeneity concerns for the specication of the CDS
The robustness checks are reported in the additional table le as an Internet Appendix. First, as shown in Table A1 for distance-to-default, we control for xed eects and nd that the coecient of CDS Active is still signicant. Second, we consider the bankruptcy prediction model used by Campbell, Hilscher, and Szilagyi (2008) and report the results in Table A2, which shows similar results. Third, we take into account the initial credit quality and the natural drift in credit quality, and show in Tables A3-A6 that our nding of the CDS eect is robust to such considerations. Fourth, as shown in Table A7, the CDS eect is stronger for non-investment grade rms. Fifth, Table A8 shows that our results are similar when we exclude rms that exit the sample as a consequence of mergers and acquisitions. 20 Recall from Figure 1 that CDS rms typically have investment grade ratings at the time of CDS introduction. Therefore, the initiation of CDS trading is not necessarily attributable to poor initial credit quality.
19

16

trading selection model. Moreover, we take into account additional considerations in choosing the explanatory variables for the CDS trading model, given that our focus is explicitly on credit risk. We employ two instrumental variables: FX hedging activities by banks and underwriters, Lender FX Usage, and the Tier One capital ratio of the lenders, Lender Tier 1 Capital.21 We rst identify lenders and bond underwriters for our sample rms based on DealScan data (for lenders) and FISD data (for bond underwriters). We then look at Federal Reserve call report data for the FX derivatives positions for these lenders and bond underwriters. For each rm in each quarter, Lender FX Usage is constructed as the average amount of FX derivatives usage for hedging purposes relative to their total assets, across banks that have either served as a lender or a bond underwriter over the previous ve years.22 To construct the instrument Lender Tier 1 Capital, we further link the identications of the lenders and bond underwriters with the Compustat Bank le containing lenders Tier One capital ratio data. For each rm in each quarter, the Lender Tier 1 Capital ratio is dened as the average of the Tier One capital ratios across banks that have either served as lenders or bond underwriters for this rm over the previous ve years.23 Besides these two instruments as explanatory variables for CDS trading, we also include rm size: Larger rms naturally attract more attention from CDS traders since the chance of hedging demand arising from any investor is greater for larger rms. In addition, we include a set of rm characteristics such as sales, tangible assets, working capital, cash holdings and capital expenditure. Furthermore, we include credit risk variables such as leverage, protability, equity volatility, and the credit rating status of the rm, for predicting the inception of CDS trading. We use data from 1997 until the rst month of CDS trading for CDS rms, and all observations for non-CDS rms, to predict the introduction of CDS trading for a rm. The prediction is estimated using a probit model: the dependent variable is equal to one after the rm starts CDS trading, and zero prior to that. The probit regression results are reported in Table III. We conrm that larger rms are more likely to have CDS contracts trading on them. CDS trading is more likely for rms with higher leverage but with investment grade ratings. Unrated rms are less likely to have CDS trading. Firms with high protability, tangibility, and large working capital are more likely to have CDS trading. Overall, it appears that rms have relatively high credit quality and visibility (a stronger balance sheet and larger size) at
Saretto and Tookes (2012) also use the rst of these IVs, Lender FX Usage, which is motivated by the ndings in Minton, Stulz, and Williamson (2009). 22 Since we are using the average FX hedging activity across all the lenders and underwriters to a rm, any selection eect that bad banks switch their lending to bad borrowers at the individual bank level would be considerably mitigated. Furthermore, this selection eect is likely to small, in any case, since the banking relationships of rms are generally stable and do not change dramatically over time. 23 The argument in the previous footnote about the mitigation of the selection eects due to the aggregation of hedging activity across all bank lenders in the case of Lender FX Usage also applies in the case of Lender Tier 1 Capital.
21

17

the time of CDS inception. Both our instrumental variables, Lender FX Usage and Lender Tier 1 Capital, are signicant predictors of CDS trading, even after controlling for other variables. Table III shows that CDS trading can be reasonably explained by the chosen variables, with pseudo-R2 s of around 38.9% across the three model specications (Models 1 and 2 include one IV at a time and Model 3 includes both IVs). In the following analysis, we will use these three CDS trading prediction models to conduct our IV estimation, treatment eects, propensity score matching, and dierence-in-dierence analyses, to re-examine the relationship between CDS trading and bankruptcy risk. We focus on the probability of bankruptcy in the remaining analysis to conserve space, although the results for the probability of a credit rating downgrade point to the same conclusion, and are available upon request.

C.1. Instrumental Variable Estimation

We rst present our IV estimation results to address the selection and endogeneity concerns. Undoubtedly, the quality of the instrumental variables is important for the consistency of such estimation results. In particular, the instruments need to satisfy the relevance and exclusion restrictions. Table III shows that CDS trading is signicantly associated with Lender FX Usage and Lender Tier 1 Capital, demonstrating their relevance to CDS trading. The exclusion restriction is impossible to test formally, as argued by Roberts and Whited (2012). The instruments we use are economically sound, because they are associated with the overall hedging interest of the lenders or credit suppliers, and their Tier One capital adequacy ratio. Moreover, the instruments we use are not weak: The F -test statistics are 56, 11, and 68 individually, and jointly they are above 10 for both IVs, which are statistically signicant. We next perform additional analysis to account for the discreteness of the CDS Active variable since the tted values of the rst stage of the two-stage least squares (2SLS) would be continuous variables. We classify CDS Active as one if the probability of having CDS trading is above the median (in the top 50%), or in the top 25% respectively.24 Table IV shows the second-stage estimation result using both Lender FX Usage and Lender Tier 1 Capital as IVs. Our instrumented CDS Active variable is signicant in all our specications. Furthermore, we run the Sargan over-identication test and cannot reject the hypothesis that both IVs are exogenous. Note that the purpose of the IV estimation is to control the endogeneity in the specic timing of CDS introduction. We next directly address the selection of rms into the CDS sample.
24

Cohen, Frazzini, and Malloy (2012) employ a similar method.

18

C.2. Heckman Treatment Eects Model

The selection of rms for CDS trading is analogous to the missing data problem in the spirit of Heckman (1979) as we do not observe the counterfactual outcome (CDS active rms without CDS trading). Therefore, correcting for self-selection can be viewed as including an omitted variable that is proxied by the Inverse Mills Ratio from the rst stage of the Heckman procedure to produce a consistent estimate.25 The selection models for CDS trading are the same probit models that underlie Table III. Based on the estimated model parameters from the rst stage, we calculate the Inverse Mills Ratio, which is a transformation of these predicted individual probabilities of CDS trading. Then, the second stage of the hazard model analysis includes the Inverse Mills Ratio as an additional explanatory variable. We include all rm observations in our second-stage analysis. The second-stage results of the Heckman correction with instrument variables are presented in Table V. We use all three CDS prediction models for the rst-stage estimation to generate the Inverse Mills Ratio. We nd that CDS Active has a positive and signicant coecient estimate in all specications. In other words, rms are more likely to go bankrupt after the introduction of CDS trading. The economic magnitude of the coecient is also large. For example, from Model 3 in the rst stage, the marginal eect of CDS trading on bankruptcy ling is 0.37%, compared with the average bankruptcy probability of 0.14% in the overall sample. Testing the signicance of the Inverse Mills Ratio is a test of whether the private information possessed by CDS traders explains the outcome, i.e., bankruptcy ling. The coecient of Inverse Mills Ratio is insignicant. These results show that the positive relationship between CDS trading and bankruptcy risk is robust to the selection of rms for CDS trading.26
We note that the Heckman model assumes a bivariate normal distribution for the error terms of the rst-stage and second-stage regressions. Thus far, there is no theory regarding alternative distributional assumptions. Theoretically, the exclusion restriction is not necessary in all applications of the Heckman selection model if the model is identied merely on account of its nonlinearity, although it is safe to impose the exclusion restriction as the selection can be approximately linear in the relevant region. Also, employing multiple instruments can be helpful if they improve the predictability of the rst stage. 26 We also tried two other instrumental variables: TRACE Coverage and Post CFMA. The pricing of CDS might be easier for rms in the TRACE (Trade Reporting and Compliance Engine) database of the Financial Industry Regulatory Authority (FINRA), due to the ease of obtaining market information in a timely manner in an OTC market. This will increase the probability of CDS trading for these rms. The Commodity Futures Modernization Act of 2000 (CFMA) ensures the deregulation of OTC derivatives. Therefore, rms are more likely to have CDS trading in the post-CFMA period. As expected, we nd these instruments to be signicant determinants of CDS trading. The CDS eect is also signicant using these instruments, although they are not our rst choices, as shown in Table A9.
25

19

C.3. Propensity Score Matching

We now re-estimate our baseline model using a propensity score matched sample. Propensity score matching makes the treatment eect easy to interpret as the dierence between the CDS rms and those without CDS traded is measured by the coecient of CDS Active. For each CDS rm, we nd one matching non-CDS rm with the nearest propensity score for CDS trading. We then run the hazard rate model on this matched sample. We use the three CDS prediction models from Table III and three dierent matching criteria: (1) the one non-CDS rm with the nearest distance, in terms of propensity score, to the CDS rm; (2) the one rm with the nearest propensity score but within a dierence of 1%, and (3) the two rms with propensity scores closest to the CDS-trading rm. We nd that there are no signicant dierences in either the propensity scores or the Z -score between the CDS rms and the matching rms, for all prediction models. Table VI presents the regression results for our CDS-trading propensity-matched sample. In all specications, the coecient estimates for CDS Active are signicantly positive. Therefore, the probability of bankruptcy increases after CDS trading has begun, even adjusting for the propensity for CDS trading. CDS Firm is not signicant in any specication. (We only present the results for the bankruptcy prediction and for the specication with the control, CDS rm, to conserve space.) Therefore, after matching by the propensity for CDS trading, CDS rms are no longer statistically signicantly dierent from non-CDS rms in terms of credit quality deterioration, attesting to the eectiveness of our matching procedure. We use CDS prediction Model 3 and the nearest one matching as a benchmark case (reported in the second column of Table VI). When we modify the matching criterion from the nearest one to the nearest one with propensity score dierence within 1%, the results are similar to those in column 2 without the 1% restriction. As an alternative, we choose two matching rms with the nearest propensity scores from Model 3, and still nd a signicant coecient estimate for CDS Active. Furthermore, we also use Models 1 and 2 with the nearest-one propensity score matching. These models produce dierent matching samples, due to the data available to calculate propensity scores for each prediction model. CDS Active is signicant in all these other specications.

C.4. Dierence-in-Dierence Analysis

Another approach that can be used to address the endogeneity concerns and identify the treatment eect (of the introduction of CDS trading) is dierence-in-dierence analysis. Similar to our propensity score matching analysis, we identify non-CDS rms matching the CDS rms using all three CDS prediction models presented in Table III and three dierent matching 20

criteria: (1) the one non-CDS rm with the nearest distance, in terms of propensity score, to the CDS rm; (2) the one rm with the nearest propensity score but within a dierence of 1%, and (3) the two rms with propensity scores closest to the CDS-trading rm in question. Furthermore, we consider three windows for the event analysis: year t 1 to year t + 1, year t 1 to year t + 2, and year t 1 to year t + 3 (where year t is the year of introduction of CDS trading). We cannot run the dierence-in-dierence analysis on the binary bankruptcy event directly. Therefore, we examine a continuous measure of probability of default: the expected default frequency (EDF ), which is a normal transformation of the distance-to-default (EDF = N (DD)). The calculation of DD follows Bharath and Shumway (2008), with an adjustment for the leverage ratio of nancial rms. There are several advantages to choosing EDF as the relevant variable to track. First, EDF is a continuous measure of credit quality. Therefore, the estimation has more power and the CDS introduction eect can be more easily identied. Second, using EDF enriches our empirical framework of credit risk measured by downgrading and bankruptcy ling. While also being a measure of credit risk, the EDF measure is sufciently dierent from rating downgrades and bankruptcy ling, as it is inferred from stock prices and balance sheet variables. Last, EDF is an ex ante measure of credit risk, while we can only observe downgrading and bankruptcy ex post. Using an alternative credit risk measure also helps demonstrate the robustness of our conclusion. Panel A of Table VII shows that the EDF dierence-in-dierence estimates are both statistically and economically signicant for the (t 1, t + 2) and (t 1, t + 3) event windows regardless of the CDS prediction model or matching criteria. For example, compared to the nearest-one propensity score matched rm from CDS Prediction Model 3, the EDF is 4.0% higher three years after CDS introduction. Recall that the average CDS rms has a BBB rating at the time of CDS introduction. Such an increase in EDF is substantial, given that the average BBB (BB) U.S. rms 3-year default probability was about 1.2% (5.4%), from 1981 to 2011, according to Standard & Poors (2012). The dierence-in-dierence estimates are insignicant for event window (t 1, t + 1), except when we use two matching rms. Therefore, the decline in credit quality after beginning CDS trading is rather gradual: there is little noticeable eect in the rst year, but the eect is signicant thereafter. In Panel B of Table VII, we nd that the leverage ratios of the reference rms also increase signicantly after CDS introduction. In the dierence-in-dierence estimation using CDS Prediction Model 3 for the matching and event window (t 1, t + 2), leverage increases by between 1.0% and 1.2% after CDS introduction. Our nding regarding the magnitude of the change in leverage following the instigation of CDS trading is consistent with the conclusions of Saretto and Tookes (2012). Further, the leverage reaction seems more rapid: the leverage increases occur mostly in the rst year after CDS trading begins. 21

C.5. Falsication Test

We have considered the appropriate approaches to addressing selection and endogeneity concerns suggested by the literature.27 Since CDS are traded over-the-counter, there could be measurement error resulting from the (unobservable) exact date of CDS introduction. Such a measurement error may lead to an attenuation bias, although this may not always be the case. We further conduct a falsication test as suggested by Roberts and Whited (2012). When we shift forward the CDS introduction by one year, the eect of CDS Active becomes insignicant, as shown in Table A10. This nding demonstrates the importance of the correct identication of the timing of CDS introduction, as well as the eect of CDS trading. Therefore, our falsication test of shifting the year of CDS introduction by one year suggests that the measurement error would indeed attenuate or even eliminate our results.

D. Eect of Outstanding CDS Positions

Our analysis so far has focused on the CDS introduction eect captured by a binary variable, CDS Active, which is a permanent regime variable. That is, once CDS trading is initiated for a rm, it cannot go back to being a non-CDS rm. Such a regime variable ignores much of the information in the variation in CDS trading over time. Indeed, CDS trading activity varies considerably over time and across rms. Such variations may generate additional implications for the eects of CDS trading on credit risk. If the CDS trading activity of a rm is very thin or illiquid, the corresponding CDS eect may be less pronounced. Also, a larger outstanding position in CDS may generate greater monetary consequences for CDS traders. Intuitively, for instance, if CDS trading causes credit risk changes, the inuence of CDS on credit risk should disappear when all outstanding CDS contracts mature and are extinguished. In this subsection, we provide a stronger test for such a symmetric and continuous eect of CDS trading. A unique advantage of our CDS transactions database is that it includes details about the notional amount of the CDS contracts outstanding and the contractual specications of each contract. Such detailed information is useful for forming other measures of CDS trading. As pointed out by Li and Prabhala (2007), the magnitude of the selection variable (i.e., quantity of CDS trading or amount outstanding) introduces an independent source of variation and helps the identication of the treatment eect, while ameliorating the selection concern. We use CDS transaction records to measure outstanding CDS contracts (similar to cumuWe also considered the BBB/BB boundary for the separation between investment and speculative grades in the spirit of a regression discontinuity. Although we do not present a detailed economic model for how this boundary, and its clientele eects among investors, aects CDS trading, the results in Table A7 show that the eect of CDS trading is more pronounced for speculative grade rms.
27

22

lative trading volume) in our sample. We use the Number of Live CDS Contracts, measured by the number of CDS contracts initiated but yet to mature as of the observation month, as a measure of open interest.28 This variable measures the breadth and consistency of CDS trading activity. CDS exposure computed as the Number of Live CDS Contracts may go up or down as and when new CDS contracts are created or old contracts mature. Therefore, this continuous measure is not as strongly aected by the selection issue analyzed at length in Section IV.C. We conjecture that CDS eects will be stronger for rms with greater amounts of CDS outstanding. We estimate the hazard model of bankruptcy ling using the CDS exposure measure instead of the indicator variable CDS Active. Table VIII reports our estimation results. We set outstanding CDS positions at zero for all non-CDS rms and include both CDS and non-CDS rms in Specication 1 of Table VIII. The estimation result shows that bankruptcy risk increases with the number of live CDS contracts, evidenced by the signicant positive coecient estimate for the Number of Live CDS Contracts. The marginal eect of this variable on the probability of bankruptcy is 0.03%. That is, when the number of CDS contracts outstanding increases by 33, its probability of default increases by 1%. The pseudo-R2 for Specication 1 of Table VIII is lower than in the previous analysis using the variable CDS Active. It is possible that the aggregate continuous variable, Number of Live CDS Contracts, is a noisy measure of the incentives of individual creditors, who may be over-insured. Moreover, the incentive eects implied by the size of the CDS position may be concave: They may atten out when the outstanding amount of CDS reaches a certain level. In Specication 2 of Table VIII, we only include CDS rms for the bankruptcy prediction using Number of Live CDS Contracts. The result shows that, the greater is the Number of Live CDS Contracts, the higher is the probability of bankruptcy. Therefore, even within the CDS sample, the number of CDS contracts outstanding plays a role in determining bankruptcy risk. In summary, a larger amount of CDS contracts outstanding is associated with a higher probability of rm bankruptcy.

E. The Mechanisms for the Eect of CDS on Credit Risk

Previous analysis shows a robust relation between CDS trading and the credit risk of the reference rms. In this subsection, we examine several mechanisms channeling the eect of CDS trading towards an increase in credit risk. There are two broad channels through which a rms bankruptcy risk could increase. The rst way is through a higher chance of getting into
Since CDS contracts are dened by their maturity, rather than their maturity date, new contracts are potentially created each trading day, depending on the level of trading activity.
28

23

nancial distress. The second is through a lower chance of getting out of nancial distress, leaving bankruptcy as the more likely outcome. CDS contracts can aect a rms credit risk in both of these ways. Firms can slip into nancial distress more easily when there is CDS trading if they take on more debt, increase their asset risk, become less protable, or have more pro-cyclical cash ows (i.e., have higher downside risk correlated with market conditions). Indeed, both our result in Panel B of Table VII and the ndings in Saretto and Tookes (2012) show that rm leverage increases following the inception of CDS trading. The increase in leverage naturally leads to an increase in credit risk. Therefore, we control for rm leverage in our regressions, both before and after the introduction of CDS, and focus on other mechanisms. CDS can reduce protability if there is negative feedback from the CDS market to the product market. In such a case, a negative shock, even though it could be pure noise, would reduce the sales and prots of the rm. This feedback eect could be used by market manipulators to accentuate the eect of the shock. Firm performance can become more correlated with the CDS market if the CDS market transmits negative information to market participants. This type of information mechanism is especially harmful during downturns. However, CDS can also reduce the information available about rms if lenders reduce monitoring and produce less information about the borrowers when their exposure to borrower default is hedged with CDS. Before we discuss these fundamental mechanisms, we study two other mechanisms that reduce the chance of successful debt workouts for rms in nancial distress. The rst is that lenders can be tougher once they are protected by CDS. The second is that creditor coordination is more dicult when there is CDS trading since their interests may not be aligned.

E.1. Tough Creditors Opposing Restructuring

The rst mechanism besides leverage that we investigate is due to tougher creditors, as in the Bolton and Oehmke (2011) model for empty creditors. That is, creditors insured with CDS protection will be tougher in the renegotiation of existing debt obligations, and consequently restructuring will be less successful, as shown in the illustrative example presented in Appendix A.29 The driver of the empty creditor mechanism is the extent of over-insurance by lenders using CDS contracts. This over-insurance with CDS directly drives the lenders incentive to force borrowers into bankruptcy by rejecting restructuring proposals, precipitating a default event and therefore receiving payments from CDS sellers. The greater the
The Trust Indenture Act of 1939 prohibits public debt restructuring without unanimous consent. Hence, public debt restructuring usually takes the form of exchange oers. As a consequence, there could be a potential holdout problem, since some bondholders may not participate in the oer. In this context, James (1996) shows that bank debt forgiveness is important for the success of public debt exchange oers.
29

24

degree of over-insurance by the empty creditor, the larger will be the benet from rejecting a restructure and potentially triggering bankruptcy. Our data do not reveal the identity of individual CDS traders. Hence, we cannot directly observe the presence of individual empty creditors or their portfolio positions. Consequently, we have to make do with aggregate proxies for the inception of CDS trading as a (noisy) proxy for the potential inuence of empty creditors. If we make the assumption that the presence of CDS implies a higher probability of empty creditors than among non-CDS rms, then our baseline nding is consistent with the empty creditor prediction. We calculate the ratio of the notional dollar amount of CDS contracts outstanding to the total dollar amount of debt outstanding at the same time, CDS Notional Outstanding/Total Debt.30 We scale the CDS position by total debt to relate the dollar amount of CDS outstanding to creditors exposure. CDS Notional Outstanding/Total Debt is a somewhat more informative, but still noisy, measure of the extent of the empty creditor concern. We emphasize that we do not need all creditors to become empty creditors for the empty creditor mechanism to manifest itself; it may take just a few or even one large empty creditor to holdout a restructuring proposal. We conjecture that bankruptcy risk is higher when CDS Notional Outstanding/Total Debt is larger. The estimation results, reported in Table IX, are consistent with the conjecture: A larger dollar amount of CDS contracts outstanding relative to rms debt outstanding is associated with a higher probability of rm bankruptcy. Empty creditors will clearly prefer rms to declare bankruptcy rather than have the rms debt restructured only if bankruptcy, but not restructuring, triggers a credit event for CDS contracts and generates payments to CDS buyers. Empty creditors will not have this incentive to the same degree if their CDS contracts also cover restructuring as a credit event. Thus, the strength of the empty creditor mechanism depends crucially on the denition of the restructuring clause in the CDS contract. We investigate the eect of dierences in contractual terms on the credit risk consequences of CDS trading. Appendix B describes the restructuring clauses in CDS contracts and their historical evolution. Essentially, there are four types of CDS contract, based on the denition of credit events: full restructuring (FR), modied restructuring (MR), modied-modied restructuring (MMR), and no restructuring (NR). For FR contracts, any type of restructuring qualies as a trigger event, and any debt obligation with a maturity of up to 30 years can be delivered in that event. Under MR also, any restructuring is included as a credit event; however, the deliverable obligations are limited to those with maturities within 30 months of the CDS contracts maturity. For MMR contracts, the deliverable obligations are relaxed to
The maximum value for CDS Notional Outstanding/Total Debt is 4.14, which is suggestive of overinsurance for such rms and the potential presence of empty creditors. (The mean is 0.10 and the median is 0.02.)
30

25

include those with maturities within 60 months of the CDS contracts maturity for restructured debt, and 30 months for other obligations. Under NR, restructuring is excluded as a credit event. Firms with more NR contracts are more subject to the empty creditor threat than those with other types of CDS. FR contracts would not be as strongly inuenced by the empty creditor incentives, as illustrated by the analysis in Appendix A.31 Figure 2 plots the number of contracts of each type in each year as observed in our CDS transaction records. The majority of rms in our sample have the MR type of clause in their CDS contracts. Types FR and MMR have a negligible presence in our sample, which is quite representative of the market as a whole, although there could be some variation at the rm level. The gure shows that there were hardly any NR CDS contracts prior to 2002. Packer and Zhu (2005) show that, in their sample period, MR contracts were just slightly more expensive than NR contracts. In such circumstances, CDS buyers would probably buy MR contracts rather than NR contracts. The proportion of CDS contracts with NR specications has increased dramatically in recent years, especially in 2007. The median (mean) fraction of NR contracts out of all CDS contracts for a reference entity is 0.61 (0.55). We also nd that there is wide variation across rms in terms of the proportion of NR contracts. One may be concerned with the endogeneity in choice of contract type, i.e., CDS buyers expecting bankruptcy to be more likely than restructuring will buy CDS contracts covering bankruptcy only. However, such endogeneity would have the same implication as when holders of NR CDS contracts have a clear preference for bankruptcy. We account for the dierences in contractual specications in the estimations reported in Table X, which include variables measuring the type of CDS contract. No Restructuring CDS Proportion is the fraction of CDS contracts with NR clauses out of all CDS contracts on the same reference entity. (This measure would be zero for rms without CDS.) Similarly, Modied Restructuring CDS Proportion is the fraction of CDS with MR clauses out of all contracts on the same reference entity. Since there are very few contracts with the FR or MMR specication in our sample, we focus only on the MR and NR types. We run separate regressions with the two CDS-type variables (reported in Specication 1 and Specication 2), and also a combined one with both of them (Specication 3). The results in Table X show that only for NR contracts do we nd a signicant positive relationship of CDS trading with bankruptcy risk, while the coecient of the MR type is not statistically signicant. The marginal eect of the No Restructuring CDS Proportion variable in the combined regression
31 Another related issue is the type of settlement. In the past, most CDS contracts were settled by physical delivery (CDS buyers delivered bonds to sellers to receive the face value). More recently, cash settlement has been the norm (CDS sellers pay the dierence between the face value and its recovery value directly to CDS buyers). Contracts settled by physical delivery may have an additional inuence from the empty creditor problem, since they may cause a squeeze in the bond market. In addition, physical delivery confers an additional cheapest to deliver option on the CDS protection buyer. Unfortunately, however, we do not have data on the delivery method.

26

on the probability of bankruptcy is 0.22% in Specication 3: the default probability of a rm with all NR CDS is 0.22% higher than that of a rm with no NR CDS. This magnitude is large in comparison to the overall sample default probability of 0.14%. We include year dummies in our regressions to control for potential time series patterns in the composition of CDS contract types.32 We nd that the regressions reported in Table X have higher pseudo-R2 s than those in Table IX, suggesting that the specication with restructuring information relating to the contracts ts the data better. Therefore, the eect of CDS Active seems to be driven by the CDS contracts with NR clauses. This nding on restructuring will likely be relevant to many more reference names in the future as more and more corporate CDS contracts use NR as the credit event specication (e.g., all CDS index constituents of the North America investment grade index CDX.NA.IG), especially after the CDS Big Bang in 2009. The results on CDS Notional Outstanding/Total Debt and No Restructuring CDS Proportion are consistent with the empty creditor model. Therefore, one mechanism for the CDS eect on credit risk is due to creditors becoming tougher in debt renegotiation, and consequently causing rms to le for bankruptcy. We note two caveats. First, empty creditors are only part of the market. CDS can be traded by any buyer and seller pair, and not just by the current creditors. CDS trading by parties unrelated to the reference rms would weaken the empty creditor mechanism and make it less likely for us to nd signicant CDS trading eects. Second, not all empty creditors can successfully force the borrower into bankruptcy.

E.2. Creditor Coordination Failure

Besides tough creditors causing bankruptcy on individual bases, creditor coordination is another important consideration for debt workout. If rms borrow money from a larger number of lenders after the inception of CDS trading, creditor coordination will be more dicult and bankruptcy more likely. Lead banks will probably not want to appear to drive their borrowers into bankruptcy, as the long-run reputational damage may outweigh the short-run gains from empty creditor trading prots. However, other lenders such as hedge funds or private equity players, who are not similarly constrained, may take advantage of CDS trading more intensively. Therefore, CDS trading may aect the size and composition of lenders to a rm. We investigate the impact of CDS introduction on the creditor relationships of a rm. The overall creditor relationship is represented in our analysis by the lending relationships available
We also segmented the sample by time, to test for the secular evolution of contract terms. We expected that the restructuring concern should have been less material in inuencing credit risk prior to 2000, when restructuring was normally included as a credit event in CDS contracts. In results not reported here, we nd that the CDS trading eect is indeed signicant only in more recent years.
32

27

from DealScan LPC data.33 For each rm in a given month, we examine the prior ve-year period for any syndicated loan facilities for this rm. Summing over all such active facilities, we compute the number of unique banks lending to the rm. Number of Banks is the change in the number of bank relationships from one year before the inception of CDS trading to two years after the inception of CDS trading. First, from a univariate dierence-in-dierence analysis, we nd that the number of bank relationships of a rm increases signicantly by 1.4, one year after the inception of CDS trading, and by 3, two years after CDS trading, relative to rms matched using the CDS trading prediction models discussed in Section IV.C. Second, we regress Number of Banks from the year before to two years after CDS trading started on a set of rm characteristics, and the CDS Active variable for CDS rms only. These event study results are reported in Panel A of Table XI. We nd that CDS trading signicantly increases the number of lenders that a rm has. On average, rms have 2.4 more lenders two years after CDS introduction, controlling for other factors that may aect the number of lenders, such as rm size and leverage. The relationship between the number of lenders and bankruptcy risk has previously been documented by, among others, Gilson, John, and Lang (1990) and Brunner and Krahnen (2008). We present similar evidence from our sample, also including the eect of CDS trading, in Panel B of Table XI. We include the Number of Banks as an additional explanatory variable in the hazard model of the rms probability of bankruptcy. The results indicate that a rms bankruptcy risk increases with the number of banking relationships, even after controlling for the direct impact of CDS trading. Therefore, the results in Table XI support Hypothesis 4 that CDS trading increases the number of creditors, which, in turn, increases bankruptcy risk.

E.3. Other Mechanisms and Further Discussion

Besides leverage, tough creditors and coordination failure, another potential channel for the CDS trading eect is via the feedback from CDS pricing. On the one hand, if CDS spreads are too high relative to the corresponding bond yield spreads, this may feed back to the rms bond market through arbitrage between the two markets, making it more costly and dicult for the rm to renance its obligations. In turn, this may cause the operating environment to worsen, leading to a deterioration of the rms credit quality.34 High CDS spreads also increase the cost of buying CDS protection, and hence reduce the incentive of creditors to become empty creditors and deter potential market manipulation. If, on the other hand, CDS spreads are underpriced or too low, informed traders have a greater incentive to buy CDS
The construction of the dataset is detailed by Chava and Roberts (2008). We thank Michael Roberts for providing the DealScan-Compustat linking le. 34 See, A Market Backres and Investors Pay, by Henry Sender, Wall Street Journal, December 5, 2002.
33

28

contracts and expect to make prots from the subsequent increase in CDS spreads.35 Our last consideration of mechanisms is the information content of CDS trading. CDS provide traders with a relatively simple instrument for going short a rms credit. The CDS market can provide information about a rms credit quality, especially its downside risk. Therefore, it is possible that some rms become riskier after CDS trading as information is impounded into prices more quickly, perhaps causing higher equity, bond, and asset volatility. This information channel could be consistent with our nding of higher credit risk after CDS trading.36 In summary, we do not nd evidence for the feedback and information channels of CDS trading eects. On the other hand, we nd strong evidence for the leverage, tougher creditor and coordination failure channels. Therefore, while we are less certain about whether CDS lead rms into nancial distress, our evidence is relatively clear that CDS increase the chances of bankruptcy compared with restructuring, for nancially distressed rms.

V.

Concluding Remarks

We nd strong evidence that the bankruptcy risk of reference rms increases after the inception of CDS trading, using a comprehensive dataset of North American corporate CDS transaction records over the period 1997-2009. This eect of CDS trading on credit risk is economically large: The odds of bankruptcy more than double after CDS trading begins for average rms. This nding is robust to the selection and endogeneity in CDS trading, using the lenders FX hedging and Tier One capital ratio as instrumental variables. We also nd that the the eect of CDS trading is related to the amount of CDS outstanding. Therefore, the bankruptcy risk of rms increases when CDS positions accumulate, and decreases when CDS contracts expire. The eect of CDS on bankruptcy risk is more pronounced when CDS payments do not cover restructuring. Moreover, the number of lenders increases after CDS trading begins, exacerbating the problems of creditor coordination. This study uncovers a real consequence of CDS trading and contributes to the ongoing debate on this important derivative market. We emphasize that, although according to our ndings rms become more vulnerable to bankruptcy once CDS start trading on them, this does not imply that CDS trading necessarily reduces social welfare. Indeed, CDS can increase
In Table A11, we nd that the eect of CDS trading on bankruptcy risk is signicant for both rms with CDS that is likely overpriced and those for which it is underpriced (as predicted by the basis between the CDS and bond yield spreads). Moreover, there is no statistically signicant dierence between these two groups. 36 We split our sample by analyst coverage in Table A12, and nd a signicant CDS eect for rms with both high and low analyst coverage. The eect is not statistically dierent between those sub-samples, suggesting that the CDS eect is not related to the information environment in which a rm operates.
35

29

debt capacity, and many previously unqualied projects may get funded due to the possibility of credit risk mitigation aorded by the CDS. Therefore, the cost associated with an increase in bankruptcy risk could be oset by the benets of an enlarged credit supply. Future work could examine the tradeo between the increased debt capacity and the bankruptcy vulnerability caused by CDS, shedding light on the overall impact of CDS trading on allocative eciency.

30

Appendices A Illustration of CDS Eects on Bankruptcy Risk

We use a simple example of a reduced-form nature to illustrate how CDS trading by creditors aects the likelihood of bankruptcy. The example is intended to convey the basic intuition of the incentives of creditors with CDS positions, and is based on the model of Bolton and Oehmke (2011). First, consider the case where there is no CDS traded on a rm. Assume that creditors lend X to the rm. If the rm is in nancial distress and consequently declares bankruptcy, creditors will recover r X, where r is the recovery rate in bankruptcy. Consider, on the other hand, that the creditors allow the rm to restructure the debt, since the recovery value of the assets in bankruptcy is less than its value as a going concern. Suppose the rm oers the creditors part of the dierence between the going concern value and the recovery value of the assets in bankruptcy, and agrees to pay them say R X, with R > r. Clearly, the creditors would consider such a restructuring favorably, and try to avoid bankruptcy.37 In general, restructuring would dominate bankruptcy. Suppose next that the creditors can also buy CDS protection against the rms credit events. Clearly, bankruptcy would always be dened as a credit event. However, restructuring may or may not be dened as a credit event, as per the clauses of the CDS contract. If restructuring is included as a credit event, we call the contract a full restructuring (FR) CDS. If it is not, we call it a no restructuring (NR) CDS.38 In the case of FR CDS, assume that the CDS premium (price) is F , in present value terms, at the time of default and that the creditors buy CDS against Y of notional value of the CDS. If the rm defaults, the creditors total payo with CDS protection is [r X + (1 r F ) Y ] in the event of bankruptcy, and [R X + (1 R F ) Y ] if the debt is restructured. Therefore, the creditors are better o with bankruptcy than with restructuring if [r X + (1 r F ) Y ] > [R X + (1 R F ) Y ], i.e., when Y > X, since R > r. Hence, bankruptcy dominates restructuring as a choice for creditors for whom the amount of CDS purchased exceeds the bonds held (empty creditors), even when restructuring is covered by the CDS. In the equilibrium model of Bolton and
The precise size of R would be determined in a bargaining process between the creditors and the shareholders of the rm. 38 Other types of CDS contracts also exist, but are not relevant for the purpose of this simple illustration. See Appendix B for a discussion of contract clauses.
37

31

Oehmke (2011), CDS sellers fully anticipate this incentive of CDS buyers, and price it into the CDS premium. Although CDS sellers may have an incentive to bail out the reference rms (by injecting more capital as long as it is less than the CDS payout) in order not to trigger CDS payments, they cannot do so unilaterally, since the empty creditors who are the CDS buyers, and other creditors, will mostly decide the fate of the company as any new nancing would require the existing creditors approval. CDS sellers are not part of this negotiation process. Now consider the case of NR CDS. Assume that the CDS premium, in this case, is f in present value terms, where f < F . Suppose again that the creditors buy CDS against Y of notional value of the CDS. Therefore, if the rm defaults, the creditors total payo with CDS protection is [r X + (1 r f ) Y ] in the event of bankruptcy, and [R X f Y ] if the debt is restructured. Bankruptcy is a preferred outcome for the creditors if [r X + (1 r f ) Y ] > [R X f Y ], Rr X, 1r which can be true even when Y < X, since R < 1. Thus, for NR CDS, bankruptcy is preferred when even a relatively small amount of CDS are purchased; hence, bankruptcy is the preferred outcome for a larger range of holdings of NR CDS by the creditors. It is also evident that buying CDS protection with NR CDS contracts is more protable in bankruptcy than restructuring without CDS protection, so long as Y > [r X + (1 r f ) Y ] > R X, which is equivalent to saying that
39

or when

Y >

Rr X. 1rf

The above condition is met when Y > X, as long as R < 1 f , which is almost always true as the cost of CDS protection is usually lower than the loss in the event of restructuring. Even if Y < X, the condition is likely to hold, for reasonable values of R and f . Further, the greater the dierence between Y and X, the greater will be the incentive for creditors to push the rm into bankruptcy. Our parsimonious illustration skips many details of the equilibrium model of Bolton and Oehmke (2011) in order to capture the main intuition and predictions. We refer interested
The calculation for the FR CDS is the same, except that the fee is replaced by F instead of f . The precise range of values for Y relative to X would be smaller than for the NR CDS, as argued above.
39

32

readers to Bolton and Oehmkes (2011) theory for a more rigorous treatment. To recap, we demonstrate that a) creditors have an incentive to over-insure and push the rm into bankruptcy, b) this incentive increases with the dierence between Y and X, i.e., the amount of CDS contracts outstanding relative to the rms debt, and c) the probability of bankruptcy occurring is greater for NR CDS contracts.

Credit Default Swaps Credit Event Denitions

Credit default swaps (CDS) provide insurance protection against the default of a reference entitys debt. For the buyer of protection to obtain payment from a CDS contract, a credit event must be triggered. Following such an event, the CDS contract can be settled either by physical delivery (by delivering the reference security and receiving the notional principal) or payment of cash (by receiving the dierence between the notional principal and the price of the reference security). The trade organization of participants in the derivatives market, the International Swaps and Derivatives Association (ISDA), sets the standards for the contractual terms of CDS contracts, including the denition of trigger events, the delivery and settlement process, and other details. Based on the 1999 ISDA Credit Event Denitions, there are six categories of trigger events for calling a default for dierent obligors: bankruptcy, failure to pay, obligation acceleration, obligation default, repudiation/moratorium and restructuring. For CDS linked to corporate debt, the primary trigger events are bankruptcy, failure to pay and restructuring. Under this denition, known as full restructuring (FR), any restructuring qualies as a trigger event, and any obligations with a maturity up to 30 years can be delivered. This creates a cheapest to deliver option for protection buyers who will benet by delivering the least expensive instrument in the event of default. The broad denition of deliverable obligations was intended to create a standard hedge contract with a wide range of protection possibilities for the credit risk of the reference entity. However, the restructuring of Conseco Finance on 22 September 2000 highlighted the problems with the 1999 ISDA Credit Event Denitions. The bank debt of Conseco Finance was restructured to the benet of the debt holders. Yet, the restructuring event still triggered payments from outstanding CDS contracts. To settle the CDS position, CDS holders also utilized the cheapest-to-deliver option created by the broad denition of deliverable obligations and delivered long-maturity, deeply discounted bonds in exchange for the notional amount. To address this obvious lacuna, ISDA modied CDS contracts and dened a new structure known as modied restructuring (MR). Under this 2001 ISDA Supplement Denition, any restructuring is dened as a credit event. However, the deliverable obligations are limited to 33

those with maturities within 30 months of the CDS contracts maturity. In March 2003, ISDA made another change and introduced modied-modied restructuring contracts (MMR) to relax the limitation on deliverable obligations. The deliverable obligations were relaxed to those with maturities within 60 months of the CDS contracts maturity for restructured debt, and 30 months for other obligations. Thus, following the 2003 ISDA Credit Derivative Denitions, there are four types of restructuring clauses: full restructuring (FR), modied restructuring (MR), modied-modied restructuring (MMR) and no restructuring (NR). For CDS contracts with NR as the restructuring clause, restructuring is excluded as a credit event: the credit event has to be either bankruptcy or the failure to pay. To further standardize the CDS market, since April 2009, ISDA has not included restructuring as a credit event for North American CDS contracts. To sum up, based on the 2003 ISDA Credit Derivative Denitions, there are four types of restructuring clauses: FR, MR, MMR and NR. The credit event in all cases includes bankruptcy and failure to pay. For CDS contracts under FR, the event also includes restructuring. Under NR, restructuring is excluded as a credit event. The other types include restructuring as a credit event, but dier in terms of the maturity of the deliverable obligations, MR being more restrictive than MMR. By 2009, the rules essentially excluded restructuring as a credit event for all North American corporate CDS contracts.

34

References
Acharya, Viral V., and Timothy C. Johnson, 2007, Insider trading in credit derivatives, Journal of Financial Economics 84, 110-141. Altman, Edward I., 1968, Financial ratios, discriminant analysis and the prediction of corporate bankruptcy, Journal of Finance 23, 589-609. Allen, Franklin, and Elena Carletti, 2006, Credit risk transfer and contagion, Journal of Monetary Economics 53, 89-111. Ashcraft, Adam B., and Joo A.C. Santos, 2009, Has the CDS market lowered the cost of a corporate debt?, Journal of Monetary Economics 56, 514-523. Arping, Stefan, 2004, Credit protection and lending relationships, Working paper, University of Amsterdam. Bedendo, Mascia, Lara Cathcart, and Lina El-Jahel, 2012, In- and out-of-court debt restructuring in the presence of credit default swaps, Working paper, Universitas Bocconi and Empirical College Business School. Bharath, Sreedhar T., and Tyler Shumway, 2008, Forecasting default with the Merton distance to default model, Review of Financial Studies 21, 1339-1369. Bharath, Sreedhar T., Venky Panchapegesan, and Ingrid M. Werner, 2010, The changing nature of chapter 11, Working paper, Arizona State University, The Goldman Sachs Group, Inc. and The Ohio State University. Blanco, Roberto, Simon Brennan, and Ian W. Marsh, 2005, An empirical analysis of the dynamic relationship between investment grade bonds and credit default swaps, Journal of Finance 60, 2255-2281. Bolton, Patrick, and Martin Oehmke, 2011, Credit default swaps and the empty creditor problem, Review of Financial Studies 24, 2617-2655. Brunner, Antje, and Jan Pieter Krahnen, 2008, Multiple lenders and corporate distress: Evidence on debt restucturing, Review of Economic Studies 75, 415-442. Campbell, John Y., Jens Hilscher, and Jan Szilagyi, 2008, In search of distress risk, Journal of Finance 63, 2899-2939. Campello, Murillo, and Rafael A. Matta, 2012, Credit default swaps, rm nancing and the economy, Working paper, Cornell University and University of Illinois. 35

Chava, Sudheer, and Robert A. Jarrow, 2004, Bankruptcy prediction with industry eects, Review of Finance 8, 537-569. Chava, Sudheer, and Michael Roberts, 2008, How does nancing impact investment? The role of debt covenants, Journal of Finance 63, 2085-2121. Che, Yeon-Koo, and Rajiv Sethi, 2012, Credit derivatives and the cost of capital, Working paper, Columbia University. Cohen, Lauren, Andrea Frazzini, and Christopher Malloy, 2012, Hiring cheerleaders: Board appointments of independent directors, Management Science, forthcoming. Damodaran, Aswath, and Marti G. Subrahmanyam, 1992, The eects of derivative securities on the markets for the underlying assets in the United States: A survey, Financial Markets, Institutions and Instruments 1, 1-22. Danis, Andras, 2012, Do empty creditors matter? Evidence from distressed exchange oers, Working paper, Vienna Graduate School of Finance. Duee, Gregory R., and Chunsheng Zhou, 2001, Credit derivatives in banking: Useful tools for managing risk?, Journal of Monetary Economics 48, 25-54. Due, Darrell, 2007, Innovations in credit risk transfer: Implications for nancial stability, Working paper, Stanford University. Gilson, Stuart C., Kose John, and Larry H. P. Lang, 1990, Troubled debt restructurings: An empirical study of private reorganization of rms in default, Journal of Financial Economics 27, 315-354. Gopalan, Radhakrishnan, Vikram Nanda, and Vijay Yerramilli, 2011, Does poor performance damage the reputation of nancial intermediaries? Evidence from the loan syndication market, Journal of Finance 66, 2083-2120. Grin, John, and Dragon Yongjun Tang, 2012, Did subjectivity play a role in CDO credit ratings?, Journal of Finance 67, 1293-1328. Hale, Galina, and Joo A.C. Santos, 2009, Do banks price their informational monopoly?, a Journal of Financial Economics 93, 185-206. Heckman, James, 1979, Sample selection bias as a specication error, Econometrica 47, 153-161. Hirtle, Beverly, 2009, Credit derivatives and bank credit supply, Journal of Financial Intermediation 18, 125-150. 36

Hu, Henry T.C., and Bernard S. Black, 2008, Debt, equity, and hybrid decoupling: Governance and systemic risk implications, European Financial Management 14, 663-709. James, Christopher, 1996, Bank debt restructurings and the composition of exchange oers in nancial distress, Journal of Finance 51, 711-727. Jarrow, Robert A., 2011, The Economics of Credit Default Swaps, Annual Review of Financial Economics 3, 235-257. Li, Kai, and Nagpurnanand R. Prabhala, 2007, Self-selection models in corporate nance, in B. Espen Eckbo ed., Handbook of Corporate Finance: Empirical Corporate Finance I, Elsevier, Chapter 2, 37-86. Longsta, Francis A., Kay Giesecke, Stephen Schaefer, and Ilya Strebulaev, 2011, Corporate bond default risk: A 150-year perspective, Journal of Financial Economics 102, 233-250. Longsta, Francis A., Sanjay Mithal, and Eric Neis, 2005, Corporate yield spreads: Default risk or liquidity? New evidence from the credit default swap market, Journal of Finance 60, 2213-2253. Merton, Robert C., 1974, On the pricing of corporate debt: the risk structure of interest rates, Journal of Finance 29, 449-470. Minton, Bernadette A., Rene M. Stulz, and Rohan Williamson, 2009, How much do banks use credit derivatives to hedge loans?, Journal of Financial Services Research 35, 1-31. Moodys, 2009, Moodys approach to evaluating distressed exchange. Morrison, Alan D., 2005, Credit derivatives, disintermediation, and investment decisions, Journal of Business 78, 621-47. Narayanan, Rajesh, and Cihan Uzmanoglu, 2012, Empty creditors and distressed debt restructuring, Working paper, Louisiana State University. Nashikkar, Amrut, Marti G. Subrahmanyam, and Sriketan Mahanti, 2011, Liquidity and arbitrage in the market for credit risk, Journal of Financial and Quantitative Analysis 46, 627-656. Norden, Lars, Consuelo Silva Buston, and Wolf Wagner, 2012, Banks use of credit derivatives and the pricing of loans: What is the channel and does it persist under adverse economic conditions?, Working paper, Erasmus University Rotterdam and Tilburg University. Packer, Frank, and Haibin Zhu, 2005, Contractual terms and CDS pricing, BIS Quarterly Review, 89-100. 37

Parlour, Christine A., and Andrew Winton, 2012, Laying o credit risk: Loan sales versus credit default swaps, Journal of Financial Economics, forthcoming. Peristiani, Stavros, and Vanessa Savino, 2011, Are credit default swaps associated with higher corporate defaults?, Working paper, Federal Reserve Bank of New York. Roberts, Michael R., and Toni M. Whited, 2012, Endogeneity in empirical corporate nance, in George Constantinides, Milton Harris, and Rene Stulz, eds., Handbook of the Economics of Finance 2, Elsevier, forthcoming. Saretto, Alessio, and Heather Tookes, 2012, Corporate leverage, debt maturity and credit supply: The role of credit default swaps, Working Paper, Yale School of Management. Shumway, Tyler, 2001, Forecasting bankruptcy more accurately: A simple hazard model, Journal of Business 74, 101-124. Sorescu, Sorin M., 2000, The eect of options on stock prices: 1973-1995, Journal of Finance 55, 487-514. Standard & Poors, 2012, 2011 Annual U.S. corporate default study and rating transitions. Stanton, Richard, and Nancy Wallace, 2011, The bears lair: Index credit default swaps and the subprime mortgage crisis, Review of Financial Studies 24, 3250-3280. Stulz, Rene M., 2010, Credit default swaps and the credit crisis, Journal of Economic Perspectives 24, 73-92. Tett, Gillian, 2009, Fools Gold: How the Bold Dream of a Small Tribe at J.P. Morgan Was Corrupted by Wall Street Greed and Unleashed a Catastrophe, The Free Press, New York, NY. Thompson, James, 2010, Counterparty Risk in Financial Contracts: Should the Insured Worry About the Insurer, Quarterly Journal of Economics 125, 11951252.

38

Figure 1: Rating Distribution Around the Introduction of Credit Default Swaps. This
gure plots the credit rating distributions for rms with credit default swaps (CDS), before the inception of CDS trading and two years after the inception of CDS trading. The credit ratings are taken from S&P Credit Ratings. The CDS data come from CreditTrade and the GFI Group. There are 901 rms in our sample that have CDS traded at some point during the sample period of June 1997 to April 2009.

39

Figure 2: Credit Default Swaps Restructuring Clauses by Year. This gure plots the
distribution of credit default swaps (CDS) restructuring clauses, by year, in our sample, between 1997 and 2009. The CDS data are taken from CreditTrade and the GFI Group. There are four types of contract terms related to restructuring: full restructuring (FR), modied restructuring (MR), modied-modied restructuring (MMR), and no restructuring (NR). For rms with NR in the restructuring clause, the credit events do not include restructuring, while for the other types, they do. MR and MMR contracts impose restrictions on the types of bond that can be delivered in the event of default.

40

Table I Credit Default Swaps Trading and Bankruptcies by Year

This table reports the distribution of rms, including those with credit default swaps (CDS) traded, and bankruptcy events, by year, in our sample between 1997 and 2009. The sample of all rms is drawn from Compustat, and includes all companies in the database during 1997-2009. The CDS data are taken from CreditTrade and the GFI Group. There are 901 rms in our sample that have CDS traded at some point during the sample period of June 1997 to April 2009. The bankruptcy data are obtained from New Generation Researchs Public and Major Company Database, the UCLA-LoPucki Bankruptcy Research Database (BRD), the Altman-NYU Salomon Center Bankruptcy List, the Fixed Income Securities Database (FISD) and Moodys Annual Reports on Bankruptcy and Recovery. The combined database includes all public companies that led for bankruptcy during the period; it also includes selected private rms that are deemed signicant. The rst column in the table is the year. The second column in the table shows the total number of U.S. companies included in the Compustat database. The third column shows the number of bankruptcies in the year. The fourth column reports the number of rms for which CDS trading was initiated during the year in question. The fth column presents rms with active CDS trading during each year. The last two columns report the number of CDS rms that led for bankruptcy and the number of non-CDS rms that led for bankruptcy, respectively. ( from June 1997, until April 2009) (1) Year 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 Total (2) Total # of Firms 9366 9546 9545 9163 8601 8190 7876 7560 7318 6993 6651 6223 5686 (3) # of Bankruptcies 50 92 118 158 257 225 156 86 76 49 61 121 179 1628 (4) # of New CDS Firms 22 58 55 102 172 221 93 58 73 28 9 9 1 901 (5) # of Active CDS Firms 22 72 106 196 334 547 582 593 629 533 418 375 234 (6) # of CDS Bankruptcies 0 0 0 1 8 12 5 0 5 2 1 4 22 60 (7) # of Non-CDS Bankruptcies 50 92 118 157 249 213 151 86 71 47 60 117 157 1568

41

Table II Impact of Credit Default Swaps Trading on Credit Quality

This table presents the estimates of the probabilities of credit downgrades and bankruptcy, using a logistic model in a sample including rms with credit default swaps (CDS) and all non-CDS rms. ln(E) is the logarithm of the rms market value of equity. ln(F) is the logarithm of the book value of the rms debt, where book debt is the sum of short-term debt and 50% of long-term debt. 1/E is the inverse of the rms annualized equity volatility. rit1 rmt1 is the rms excess return over the past year, and NI/TA is the rms ratio of net income to total assets. To estimate the impact of CDS trading on the probability of credit downgrades or bankruptcy, we include CDS variables in the model specication. CDS Firm equals one if the rm is in the CDS sample and zero otherwise. CDS Active is a dummy variable that equals one after the inception of CDS trading and zero before CDS trading. The coecient of interest is that of CDS Active, which captures the impact of CDS trading on the probability of credit downgrades or bankruptcy after the inception of CDS trading. The sample period is 1997-2009, based on monthly observations. (*** Signicant at 1% level, ** signicant at 5% level, and * signicant at 10% level. The numbers in parentheses are standard errors.) Probability of Downgrades (2) (1) 0.735 0.736 (0.014) (0.014) 0.507 0.503 (0.015) (0.015) 0.062 0.017 (0.027) (0.026) 0.281 0.252 (0.035) (0.035) 0.003 0.000 (0.025) (0.024) 0.755 (0.057) 0.691 1.371 (0.067) (0.045) Yes Yes Yes Yes 15.08% 14.75% 658966 658966 3863 3863 1.925 3.939 0.39% 0.78% 0.58% 0.59% Probability of Bankruptcy (4) (3) 0.713 0.710 (0.024) (0.024) 0.711 0.713 (0.023) (0.023) 1.626 1.675 (0.131) (0.131) 1.320 1.331 (0.111) (0.111) 0.038 0.038 (0.013) (0.013) 2.009 (0.711) 2.373 0.400 (0.729) (0.177) Yes Yes Yes Yes 24.18% 24.06% 658966 658966 940 940 10.730 1.492 0.33% 0.06% 0.14% 0.14%

ln(E) ln(F) 1/E rit1 rmt1 NI/TA CDS Firm CDS Active Time Fixed Eects Industry Fixed Eects Pseudo R2 N # of Downgrades (Bankruptcy) CDS Active Odds Ratio CDS Active Marginal Eect Sample Probability of a Downgrade (Bankruptcy)

42

Table III Probability of Credit Default Swaps Trading

This table presents the estimates of the probability of credit default swaps (CDS) trading, obtained using a probit model. Propensity scores are estimated based on the model parameters. ln(Assets) is the logarithm of the rms total assets value. Leverage is dened as the ratio of book debt to the sum of book debt and market equity, where book debt is the sum of short-term debt and 50% of long-term debt and market equity is the measure of the number of common shares outstanding multiplied by the stock price. ROA is the rms return on assets. rit1 rmt1 is the rms excess return over the past year. Equity Volatility is the rms annualized equity volatility. PPENT/Total Asset is the ratio of property, plant and equipment to total assets. Sales/Total Asset is the ratio of sales to total assets. EBIT/Total Asset is the ratio of earnings before interest and tax to total assets. WCAP/Total Asset is the ratio of working capital to total assets. RE/Total Asset is the ratio of retained earnings to total assets. Cash/Total Asset is the ratio of cash to total assets. CAPX/Total Asset is the ratio of capital expenditures to total assets. Investment Grade is a dummy variable that equals one if the rm has an investment grade (BBB- and above) rating. Rated is a dummy variable that equals one if the rm is rated. Lender FX Usage is a measure of the FX hedging activities by the lending banks and underwriters and Lender Tier 1 Capital is the Tier One capital ratio of the lenders. The sample period is 1997-2009, based on monthly observations. (*** Signicant at 1% level, ** signicant at 5% level, and * signicant at 10% level. The numbers in parentheses are standard errors.)

43

Ln(Assets) Leverage ROA rit1 rmt1 Equity Volatility PPENT/Total Asset Sales/Total Asset EBIT/Total Asset WCAP/Total Asset RE/Total Asset Cash/Total Asset CAPX/Total Asset Investment Grade Rated Lender FX Usage Lender Tier 1 Capital F-statistic (instruments) p-value (F-statistic) Time Fixed Eects Industry Fixed Eects Pseudo R2 N #CDS Event

CDS Prediction Model 1 0.794 (0.005) 0.401 (0.026) 0.019 (0.016) 0.100 (0.010) 0.067 (0.015) 0.349 (0.029) 0.021 (0.003) 0.249 (0.059) 0.149 (0.024) 0.020 (0.005) 0.251 (0.035) 1.833 (0.115) 0.916 (0.013) 0.957 (0.015) 2.487 (0.732)

Probability of CDS Trading CDS Prediction Model 2 0.798 (0.005) 0.409 (0.026) 0.018 (0.017) 0.099 (0.010) 0.069 (0.015) 0.358 (0.029) 0.021 (0.003) 0.261 (0.060) 0.154 (0.024) 0.020 (0.005) 0.254 (0.035) 1.861 (0.115) 0.912 (0.013) 0.963 (0.015)

56.15 0.000 Yes Yes 38.95% 690111 551

2.369 (0.713) 11.05 0.001 Yes Yes 38.78% 690111 551

CDS Prediction Model 3 0.795 (0.005) 0.400 (0.026) 0.019 (0.017) 0.100 (0.010) 0.068 (0.015) 0.350 (0.029) 0.021 (0.003) 0.250 (0.060) 0.149 (0.024) 0.020 (0.005) 0.254 (0.034) 1.826 (0.115) 0.915 (0.013) 0.957 (0.015) 5.523 (0.732) 2.458 (0.713) 68.10 0.000 Yes Yes 38.97% 690111 551

44

Table IV Credit Default Swaps Trading and Probability of Bankruptcy: Instrumental Variable Estimation
This table presents the second-stage estimation results of the instrumental variable estimation. The second-stage analysis is for the probability of bankruptcy using a logistic model in a sample including rms with credit default swaps (CDS) and all non-CDS rms. ln(E) is the logarithm of the rms market value of equity. ln(F) is the logarithm of the book value of the rms debt, where book debt is the sum of short-term debt and 50% of long-term debt. 1/E is the inverse of the rms annualized equity volatility. rit1 rmt1 is the rms excess return over the past year, and NI/TA is the rms ratio of net income to total assets. To estimate the impact of CDS trading on the probability of bankruptcy, we include CDS variables in the model specication. CDS Firm equals one if the rm is in the CDS sample and zero otherwise. We classify CDS Active as one if the probability of having CDS trading is above the median (in the top 50%), or in the top 25% respectively, the resulting variables being dened as Instrumented CDS Active. The sample period is 1997-2009, based on monthly observations. (*** Signicant at 1% level, ** signicant at 5% level, and * signicant at 10% level. The numbers in parentheses are standard errors.) Probability of Bankruptcy CDS Prediction Model 3 ln(E) ln(F) 1/E rit1 rmt1 NI/TA CDS Firm Instrumented CDS Active Time Fixed Eects Industry Fixed Eects Pseudo R2 N # of Bankruptcy CDS Active Odds Ratio CDS Active Marginal Eect Sample Probability of a Bankruptcy Top 50% 0.625 0.623 (0.023) (0.023) 0.642 0.642 (0.022) (0.022) 1.487 1.505 (0.129) (0.128) 1.334 1.336 (0.109) (0.109) 0.033 0.033 (0.013) (0.013) 0.171 (0.167) 0.302 0.294 (0.083) (0.083) Yes Yes Yes Yes 22.30% 22.29% 657438 657438 940 940 1.353 1.342 0.04% 0.04% 0.14% 0.14% Top 25% 0.623 0.622 (0.023) (0.023) 0.644 0.644 (0.022) (0.022) 1.454 1.477 (0.127) (0.126) 1.336 1.340 (0.109) (0.109) 0.033 0.033 (0.013) (0.013) 0.261 (0.172) 0.339 0.298 (0.101) (0.098) Yes Yes Yes 22.28% 22.27% 657438 657438 940 940 1.404 1.347 0.05% 0.04% 0.14% 0.14%

45

Table V Credit Default Swaps Trading and Probability of Bankruptcy: Heckman Treatment Eects Model with Instrument Variables
This table presents the second-stage estimation results of the two-stage Heckman treatment eects model. The second-stage analysis is on the probability of bankruptcy, using a logistic model in a sample including rms with credit default swaps (CDS) and all non-CDS rms. ln(E) is the logarithm of the rms market value of equity. ln(F) is the logarithm of the book value of the rms debt, where book debt is the sum of short-term debt and 50% of long-term debt. 1/E is the inverse of the rms annualized equity volatility. rit1 rmt1 is the rms excess return over the past year, and NI/TA is the rms ratio of net income to total assets. To estimate the impact of CDS trading on the probability of bankruptcy, we include CDS variables in the model specication. CDS Firm equals one if the rm is in the CDS sample and zero otherwise. CDS Active is a dummy variable which equals one after the inception of CDS trading and zero before CDS trading. The coecient of interest is that of CDS Active, which captures the impact of CDS trading on the probability of bankruptcy after the inception of CDS trading. The Inverse Mills Ratio is calculated from the rst-stage probit regression, modeling the probability of CDS trading presented in Table III. The sample period is 1997-2009, based on monthly observations. (*** Signicant at 1% level, ** signicant at 5% level, and * signicant at 10% level. The numbers in parentheses are standard errors.) Probability of Bankruptcy CDS Prediction Model 2 0.639 (0.022) 0.645 (0.022) 1.399 (0.125) 1.330 (0.109) 0.032 (0.013) 2.269 (0.710) 2.624 (0.746) 0.040 (0.123) Yes Yes 22.42% 657438 940 13.791 0.37% 0.14%

ln(E) ln(F) 1/E rit1 rmt1 NI/TA CDS Firm CDS Active Inverse Mills Ratio Time Fixed Eects Industry Fixed Eects Pseudo R2 N # of Bankruptcy CDS Active Odds Ratio CDS Active Marginal Eect Sample Probability of a Bankruptcy

CDS Prediction Model 1 0.639 (0.022) 0.645 (0.022) 1.400 (0.125) 1.330 (0.109) 0.032 (0.013) 2.270 (0.710) 2.631 (0.746) 0.035 (0.124) Yes Yes 22.42% 657438 940 13.888 0.37% 0.14%

CDS Prediction Model 3 0.639 (0.022) 0.645 (0.022) 1.400 (0.125) 1.330 (0.109) 0.032 (0.013) 2.270 (0.710) 2.630 (0.746) 0.036 (0.124) Yes Yes 22.42% 657438 940 13.874 0.37% 0.14%

46

Table VI Credit Default Swaps Trading and Credit Quality: Propensity Score Matching
This table presents the estimates of the probability of bankruptcy using a logistic model in a sample including rms with credit default swaps (CDS) and non-CDS propensity score matched rms. Propensity score matched rms are selected based on propensity scores estimated from the model of probability of CDS trading presented in Table III. ln(E) is the logarithm of the rms market value of equity. ln(F) is the logarithm of the book value of the rms debt, where book debt is the sum of short-term debt and 50% of long-term debt. 1/E is the inverse of the rms annualized equity volatility. rit1 rmt1 is the rms excess return over the past year, and NI/TA is the rms ratio of net income to total assets. To estimate the impact of CDS trading on the probability of bankruptcy, we include CDS variables in the model specication. CDS Firm equals one if the rm is in the CDS sample and zero otherwise. CDS Active is a dummy variable that equals one after the inception of CDS trading and zero before CDS trading. The coecient of interest is that of CDS Active, which captures the impact of CDS trading on the probability of bankruptcy after the inception of CDS trading. The second column presents the analysis in the baseline matched sample, i.e. the nearest one propensity score matching rms selected based on CDS prediction model 3 in Table III. The third column presents the same analysis, but for the nearest one with propensity score dierence within 1%. The fourth column uses the two matching rms with the nearest propensity scores. The last two columns present the analysis in the matched sample selected based on CDS prediction models 1 and 2 in Table III. The sample period is 1997-2009, based on monthly observations. (*** Signicant at 1% level, ** signicant at 5% level, and * signicant at 10% level. The numbers in parentheses are standard errors.) Probability of Bankruptcy CDS Prediction CDS Prediction Model 3 Model 1 Nearest One Nearest Two Nearest One PS Di<1% Matching Matching 1.005 0.869 1.152 (0.138) (0.111) (0.149) 0.918 0.881 1.183 (0.127) (0.102) (0.139) 0.029 0.309 0.013 (0.295) (0.292) (0.309) 2.104 2.738 2.427 (0.647) (0.595) (0.699) 2.478 0.041 0.001 (0.790) (0.122) (0.177) 0.979 0.797 0.425 (0.813) (0.753) (0.795) 2.215 1.935 1.583 (0.835) (0.770) (0.781) Yes Yes Yes Yes Yes Yes 32.79% 33.35% 37.25% 111331 173665 113886 48 62 45 9.161 6.924 4.870 0.09% 0.07% 0.05% 0.04% 0.04% 0.04%

ln(E) ln(F) 1/E rit1 rmt1 NI/TA CDS Firm CDS Active Time Fixed Eects Industry Fixed Eects Pseudo R2 N # of Bankruptcy CDS Odds Ratio CDS Marginal Eect Sample Probability of a Bankruptcy

Nearest One Matching 1.009 (0.133) 0.965 (0.123) 0.069 (0.295) 2.299 (0.641) 0.012 (0.190) 0.856 (0.783) 1.968 (0.796) Yes Yes 32.78% 120975 49 7.156 0.08% 0.04%

CDS Prediction Model 2 Nearest One Matching 0.989 (0.133) 0.993 (0.121) 0.163 (0.301) 2.140 (0.628) 0.006 (0.165) 0.912 (0.783) 1.947 (0.795) Yes Yes 32.14% 120494 48 7.008 0.07% 0.04%

47

Table VII Changes in EDF and Leverage Around the Introduction of Credit Default Swaps: Dierence-in-Dierence Analysis

This table presents a univariate analysis of changes in EDF and leverage from one year before to one year, two years or three years after the introduction of credit default swaps (CDS) trading. The changes in EDF and leverage of CDS-trading rms are compared with those of the matching rms. Matching rms are selected based on propensity scores estimated from the models for the probability of CDS trading presented in Table III. The change in EDF is the change in the rms expected default frequency. EDF is calculated based on the Merton (1974) model. The change in leverage is dened as the change in the ratio of book debt to the sum of book debt and market equity, where book debt is the sum of short-term debt and 50% of long-term debt, and market equity is the measure of the number of common shares outstanding multiplied by the stock price. (*** Signicant at 1% level, ** signicant at 5% level, and * signicant at 10% level.)

48
Panel B: Change in Leverage Year t-1 to t+2 Nearest Nearest Nearest One Two One PS Di <1% 0.010 0.011 0.014 0.009 0.007 0.009 0.010 0.010 0.012

CDS Prediction Model 1 CDS Prediction Model 2 CDS Prediction Model 3

Year t-1 to t+1 Nearest Nearest One One PS Di <1% 0.005 0.007 0.007 0.001 0.008 0.001 Nearest Two 0.020 0.021 0.027

Panel A: Change in EDF Year t-1 to t+2 Nearest Nearest Nearest One Two One PS Di <1% 0.007 0.017 0.006 0.015 0.020 0.010 0.016 0.024 0.014 Nearest One 0.046 0.033 0.040

Year t-1 to t+3 Nearest One PS Di <1% 0.035 0.020 0.028

Nearest Two 0.043 0.035 0.040

CDS Prediction Model 1 CDS Prediction Model 2 CDS Prediction Model 3

Year t-1 to t+1 Nearest Nearest One One PS Di <1% 0.012 0.011 0.008 0.006 0.006 0.008

Nearest Two 0.012 0.012 0.012

Nearest One 0.014 0.013 0.007

Year t-1 to t+3 Nearest One PS Di <1% 0.018 0.016 0.008

Nearest Two 0.014 0.013 0.012

Table VIII CDS Exposure and the Probability of Bankruptcy

This table investigates the impact of credit default swaps (CDS) exposure on a rms probability of bankruptcy. Model 1 conducts the analysis in a sample including rms with CDS and all non-CDS rms. Model 2 only includes rms with CDS. ln(E) is the logarithm of the rms market value of equity. ln(F) is the logarithm of the book value of the rms debt, where book debt is the sum of short-term debt and 50% of long-term debt. 1/E is the inverse of the rms annualized equity volatility. rit1 rmt1 is the rms excess return over the past year, and NI/TA is the rms ratio of net income to total assets. CDS Firm equals one if the rm has CDS trading at any point in time and zero otherwise. CDS exposure is measured as the logarithm of the number of live CDS contracts (Number of Live CDS Contracts). The sample period is 1997-2009, based on monthly observations. (*** Signicant at 1% level, ** signicant at 5% level, and * signicant at 10% level. The numbers in parentheses are standard errors.) Probability of Bankruptcy (1) (2) 0.689 0.970 (0.026) (0.167) 0.651 0.995 (0.026) (0.166) 1.535 1.163 (0.103) (0.381) 0.622 0.518 (0.075) (0.383) 0.076 0.643 (0.023) (1.541) 0.644 (0.210) 0.240 0.539 (0.077) (0.203) Yes Yes Yes Yes 15.84% 25.53% 658966 70038 940 40 1.271 1.714 0.03% 0.03% 0.14% 0.06%

ln(E) ln(F) 1/E rit1 rmt1 NI/TA CDS Firm Number of Live CDS Contracts Time Fixed Eects Industry Fixed Eects Pseudo R2 N # of Bankruptcies Number of Live CDS Contracts Odds Ratio Number of Live CDS Contracts Marginal Eect Sample Probability of a Bankruptcy

49

Table IX Empty Creditors and the Probability of Bankruptcy

This table investigates the impact of credit default swaps (CDS) on a rms probability of bankruptcy. ln(E) is the logarithm of the rms market value of equity. ln(F) is the logarithm of the book value of the rms debt, where book debt is the sum of short-term debt and 50% of long-term debt. 1/E is the inverse of the rms annualized equity volatility. rit1 rmt1 is the rms excess return over the past year, and NI/TA is the rms ratio of net income to total assets. CDS Firm equals one if the rm has CDS trading at any point in time and zero otherwise. The empty creditor problem is measured as the total notional CDS outstanding, scaled by the book value of the total debt (CDS Notional Outstanding/Total Debt). The sample period is 1997-2009, based on monthly observations. (*** Signicant at 1% level, ** signicant at 5% level, and * signicant at 10% level. The numbers in parentheses are standard errors.) Probability of Bankruptcy 0.689 (0.026) 0.652 (0.026) 1.533 (0.104) 0.620 (0.075) 0.076 (0.023) 0.582 (0.211) 0.071 (0.032) Yes Yes 15.82% 658966 940 1.074 0.01% 0.14%

ln(E) ln(F) 1/E rit1 rmt1 NI/TA CDS Firm CDS Notional Outstanding/Total Debt Time Fixed Eects Industry Fixed Eects Pseudo R2 N # of Bankruptcies CDS Notional Outstanding/Total Debt Odds Ratio CDS Notional Outstanding/Total Debt Marginal Eect Sample Probability of a Bankruptcy

50

Table X Restructuring Clauses of CDS Contracts and Probability of Bankruptcy

This table investigates the impact of the restructuring clauses of credit default swaps (CDS) on the probability of bankruptcy of rms in a sample including rms with and without CDS traded. The empty creditor problem is expected to be more signicant for rms with more contracts with no restructuring as the restructuring clause. In Model 1, for each CDS rm, we include a variable for the No Restructuring CDS Proportion, which is the total amount of active CDS contracts with no restructuring as the restructuring clause, scaled by the total number of CDS contracts trading on it. In Model 2, for each CDS rm, we also calculate the Modied Restructuring CDS Proportion, which is the total amount of active CDS contracts with modied restructuring as the restructuring clause, scaled by the total number of CDS contracts trading on it. CDS Firm equals one if the rm has CDS trading at any point in time and zero otherwise. The coecient of interest is that of No Restructuring CDS Proportion, which captures the impact of the CDS-induced empty creditor problem. ln(E) is the logarithm of the rms market value of equity. ln(F) is the logarithm of the book value of the rms debt, where book debt is the sum of short-term debt and 50% of long-term debt. 1/E is the inverse of the rms annualized equity volatility. rit1 rmt1 is the rms excess return over the past year, and NI/TA is the rms ratio of net income to total assets. The sample period is 1997-2009, based on monthly observations. (*** Signicant at 1% level, ** signicant at 5% level, and * signicant at 10% level. The numbers in parentheses are standard errors.) Probability of Bankruptcy (2) 0.717 (0.024) 0.716 (0.023) 1.645 (0.131) 1.327 (0.111) 0.037 (0.013) 0.163 (0.210)

ln(E) ln(F) 1/E rit1 rmt1 NI/TA CDS Firm No Restructuring CDS Proportion Modied Restructuring CDS Proportion Time Fixed Eects Industry Fixed Eects Pseudo R2 N # of Bankruptcy NR CDS Odds Ratio MR CDS Odds Ratio NR CDS Marginal Eect MR CDS Marginal Eect Sample Probability of a Bankruptcy

(1) 0.716 (0.024) 0.715 (0.023) 1.636 (0.132) 1.327 (0.111) 0.037 (0.013) 0.206 (0.195) 1.315 (0.565)

Yes Yes 24.06% 658966 940 3.725 0.18% 0.14%

0.572 (0.492) Yes Yes 24.04% 658966 940 1.772 0.01% 0.14%

(3) 0.716 (0.024) 0.715 (0.023) 1.641 (0.132) 1.325 (0.111) 0.037 (0.013) 0.432 (0.255) 1.557 (0.599) 0.858 (0.528) Yes Yes 24.08% 658966 940 4.745 2.358 0.22% 0.12% 0.14%

51

Table XI CDS Trading, Bank Relationships and Probability of Bankruptcy

This table shows the results of an analysis of the impact of credit default swaps (CDS) on rm-creditor relationships. The creditor relationships are measured by bank relationships obtained from Dealscan LPC. For each rm, on a given date, we look back ve years for any syndicated loan facilities extended to this rm. Summing over all such active facilities, we compute, on each date, the number of unique bank relationships. Number of Banks is the change in the number of bank relationships from one year before to two years after the inception of CDS trading. ln(Asset) is the change in the logarithm of the rms total assets value. ROA is the change in the rms return on assets. Leverage is the change in leverage. PPENT/Total Asset is the change in the ratio of property, plant and equipment to total assets. CDS Active is a dummy variable that equals one after and zero before the inception of CDS trading. ln(E) is the logarithm of the rms equity value. ln(F) is the logarithm of the book value of the rms debt, where book debt is the sum of short-term debt and 50% of long-term debt. 1/E is the inverse of the rms annualized equity volatility. rit1 rmt1 is the rms excess return over the past year, and NI/TA is the rms ratio of net income to total assets. CDS Firm equals one if the rm has CDS trading at any point in time and zero otherwise. Number of Banks is the number of existing bank relationships. The coecients of interest are those of CDS Active and Number of Banks. (*** Signicant at 1% level, ** signicant at 5% level, and * signicant at 10% level. The numbers in parentheses are standard errors.) Panel A: CDS and Bank Relationships Number of Banks ln(Asset) ROA Leverage PPENT/Total Asset CDS Active Time Fixed Eects Industry Fixed Eects R2 N 6.291 (1.849) 0.396 (2.76) 8.581 (5.201) 1.586 (10.84) 2.432 (1.069) Yes Yes 9.75% 496 Panel B: Bank Relationships and Bankruptcy Risk Probability of Bankruptcy ln(E) 0.669 (0.026) ln(F) 0.683 (0.024) 1/E 1.763 (0.136) rit1 rmt1 1.339 (0.111) NI/TA 0.040 (0.013) CDS Firm 2.210 (0.712) CDS Active 2.378 (0.728) Number of Banks 0.153 (0.035) Time Fixed Eects Industry Fixed Eects Pseudo R2 N # of Bankruptcy CDS Active Odds Ratio Number of Banks Odds Ratio CDS Active Marginal Eect Number of Banks Marginal Eect Sample Probability of Bankruptcy Yes Yes 24.32% 658966 940 10.783 1.165 0.33% 0.02% 0.14%

52

Additional Tables
Table A1 Firm Fixed Eect Regressions for Distance-to-Default and Credit Default Swaps
This table presents estimates of the eect of CDS trading on rms distance-to-default (DD). DD is calculated from the Merton (1974) model. ln(E) is the logarithm of the rms equity value. ln(F) is the logarithm of the rms debt. 1/E is the inverse of the rms annualized equity volatility. rit1 rmt1 is the rms excess return over the past year, and NI/TA is the rms ratio of net income to total assets. To estimate the impact of CDS trading on rms DD, we include CDS variables in the model specication. CDS Active is a dummy variable that equals one if the rm has CDS traded on its debt, one year before month t. The sample period is 1997-2009, based on monthly observations. The regression controls for rm xed eects and time xed eects. (*** Signicant at 1% level,** signicant at 5% level, and * signicant at 10% level. The numbers in parentheses are standard errors.) Distance-to-Default ln(E) 0.667 (0.002) ln(F) 0.644 (0.002) 1/E 1.603 (0.003) rit1 rmt1 0.099 (0.001) NI/TA 0.031 (0.002) CDS Active 0.249 (0.008) Time Fixed Eects Yes Firm Fixed Eects Yes R-Square 82.76% N 648242

53

Table A2 Impact of Credit Default Swaps Trading on Bankruptcy: Alternative Model

This table presents estimates of the eect of CDS trading on rms bankruptcy risk, based on the model in Campbell, Hilscher, and Szilagyi (2008). NIMTAAVG is the weighted average protability ratio of net income to market-valued total assets, which includes lagged information about protability, as dened in Campbell, Hilscher, and Szilagyi (2008). TLMTA is total liabilities over the market value of total assets. EXRETAVG is the weighted average excess return over the value-weighted S&P 500 return, which includes lagged information about excess returns. Sigma is the square root of the sum of squared rm stock returns over a 3-month period. Rsize is the relative size of each rm, measured as the log ratio of its market capitalization to that of the S&P 500 index, and CASHMTA is the stock of cash and short-term investments over the market value of total assets. MB is the market-to-book ratio of the rm, and PRICE is the rms log price per share, truncated above at $15. To estimate the impact of CDS trading on rms bankruptcy risk, we include CDS variables in the model specication. CDS Firm equals one if the rm is in the CDS sample and zero otherwise. CDS Active is a dummy variable that equals one if the rm has CDS traded on its debt one year before month t. The sample period is 1997-2009, based on monthly observations. The regression controls for rm xed eects and industry xed eects. (*** Signicant at 1% level,** signicant at 5% level, and * signicant at 10% level. The numbers in parentheses are standard errors.) Probability of Bankruptcy (1) (2) 18.007 17.918 (1.697) (1.695) 3.154 3.268 (0.160) (0.162) 1.272 1.273 (0.743) (0.741) 0.829 0.800 (0.131) (0.130) 0.114 0.203 (0.031) (0.034) 2.368 2.436 (0.402) (0.404) 0.001 0.001 (0.000) (0.000) 0.429 0.485 (0.071) (0.071) 2.284 (0.456) 1.749 (0.482) Yes Yes Yes Yes 12.35% 12.77% 682053 682053 888 888 5.749 0.23% 0.13% 0.13%

NIMTAAVG TLMTA EXRETAVG Sigma Rsize CASHMTA MB PRICE CDS Firm CDS Active Time Fixed Eects Industry Fixed Eects Pseudo R2 N # of Bankruptcy CDS Active Odds Ratio CDS Active Marginal Eect Sample Probability of Bankruptcy

54

Table A3 Probability of Bankruptcy Controlling for Direct Eect of Downgrade

This table investigates the impact of credit rating and credit default swaps (CDS) trading on the probability of bankruptcy. The hazard model analysis of the probability of bankruptcy is conducted in a sample including rms with CDS and non-CDS rms, matched by their propensity score. Propensity score matched rms are selected based on propensity scores estimated from the model of probability of CDS trading presented in Table III. ln(E) is the logarithm of the rms equity value. ln(F) is the logarithm of the rms debt. 1/E is the inverse of the rms annualized equity volatility. rit1 rmt1 is the rms excess return over the past year, and NI/TA is the rms ratio of net income to total assets. CDS Firm equals one if the rm is in the CDS sample and zero otherwise. CDS Active is a dummy variable that equals one if the rm has CDS traded on its debt one year before month t. Unrated equals one if there is no credit rating on the rm. Downgrade is a dummy variable that equals one if the rm was downgraded one year before month t. The sample period is 1997-2009, based on monthly observations. (*** Signicant at 1% level, ** signicant at 5% level, and * signicant at 10% level. The numbers in parentheses are standard errors.) Probability of Bankruptcy CDS Prediction Model 1 CDS Prediction Model 2 (1) (2) (3) (4) 1.130 1.141 1.005 1.022 (0.150) (0.149) (0.137) (0.136) 1.143 1.157 0.996 1.017 (0.137) (0.137) (0.123) (0.123) 0.068 0.084 0.033 0.049 (0.244) (0.234) (0.230) (0.220) 2.118 2.140 1.895 1.880 (0.674) (0.671) (0.615) (0.607) 0.054 0.046 0.060 0.052 (0.185) (0.182) (0.164) (0.163) 0.576 0.981 (0.799) (0.787) 1.656 1.176 2.107 1.264 (0.798) (0.398) (0.810) (0.388) 1.309 1.285 1.876 1.855 (0.403) (0.401) (0.368) (0.367) 1.060 1.060 1.155 1.168 (0.442) (0.443) (0.404) (0.406) Yes Yes 38.91% 113886 45 5.238 2.886 0.11% 0.036% 0.06% Yes Yes 38.83% 113886 45 3.241 2.886 0.04% 0.04% 0.04% Yes Yes 35.34% 120494 48 8.224 3.174 0.08% 0.04% 0.04% Yes Yes 35.12% 120494 48 3.540 3.216 0.05% 0.04% 0.04%

ln(E) ln(F) 1/E rit1 rmt1 NI/TA CDS Firm CDS Active Unrated Downgrade

Time Fixed Eects Industry Fixed Eects Pseudo R2 N # of Bankruptcies CDS Active Odds Ratio Downgrade Odds Ratio CDS Active Marginal Eect Downgrade Marginal Eect Sample Probability of Bankruptcy

55

Table A4 Rating Drift and the Impact of Credit Default Swaps

This table presents the estimates of the probability of bankruptcy using a logistic model in a sample including rms with credit default swaps (CDS) and non-CDS rms matched by credit rating. The matched rms selected are the one rm with the same credit rating as the target rm and the closest asset size. ln(E) is the logarithm of the rms equity value. ln(F) is the logarithm of the book value of the rms debt. 1/E is the inverse of the rms annualized equity volatility. rit1 rmt1 is the rms excess return over the past year, and NI/TA is the rms ratio of net income to total assets. To estimate the impact of CDS trading on the probability of credit downgrades/bankruptcy, we include CDS variables in the model specication. CDS Firm equals one if the rm is in the CDS sample and zero otherwise. CDS Active is a dummy variable that equals one if the rm has CDS traded on its debt one year before month t. The coecient of interest is that of CDS Active, which captures the impact of CDS trading on the probability of bankruptcy after the inception of CDS trading. The sample period is 1997-2009, based on monthly observations. (*** Signicant at 1% level, ** signicant at 5% level, and * signicant at 10% level. The numbers in parentheses are standard errors.) Probability of Bankruptcy (1) (2) ln(E) 1.552 1.562 (0.153) (0.153) ln(F) 1.449 1.463 (0.153) (0.153) 1/E 0.548 0.528 (0.300) (0.300) rit1 rmt1 0.695 0.733 (0.448) (0.448) NI/TA 4.102 4.118 (0.643) (0.639) CDS Firm 0.530 (0.779) CDS Active 2.431 2.134 (0.667) (0.465) Time Fixed Eects Yes Yes Industry Fixed Eects Yes Yes Pseudo R2 39.95% 39.91% N 141006 141006 # of Bankruptcy 65 65 CDS Odds Ratio 11.370 8.449 CDS Marginal Eect 0.10% 0.09% Sample Probability of a Bankruptcy 0.05% 0.05%

56

Table A5 Impact of Credit Default Swaps Trading on Credit Quality: Control for Distance-to-Default
This table presents the estimates of the probability of bankruptcy using a logistic model. The analysis is conducted in a sample including rms with credit default swaps (CDS) and all non-CDS rms. Besides the conventional determinants of bankruptcy risk, we also control for rms distance-to-default (DD). DD is calculated from the Merton (1974) model. ln(E) is the logarithm of the rms market value of equity. ln(F) is the logarithm of the book value of the rms debt, where book debt is the sum of short-term debt and 50% of long-term debt. 1/E is the inverse of the rms annualized equity volatility. rit1 rmt1 is the rms excess return over the past year, and NI/TA is the rms ratio of net income to total assets. To estimate the impact of CDS trading on the probability of bankruptcy, we include CDS variables in the model specication. CDS Firm equals one if the rm is in the CDS sample and zero otherwise. CDS Active is a dummy variable that equals one after and zero before the inception of CDS trading. The coecient of interest is that of CDS Active, which captures the impact of CDS trading on the probability of credit downgrades or bankruptcy after the inception of CDS trading. The sample period is 1997-2009, based on monthly observations. (*** Signicant at 1% level, ** signicant at 5% level, and * signicant at 10% level. The numbers in parentheses are standard errors.) Probability of Downgrades 0.567 (0.020) 0.321 (0.020) 0.315 (0.035) 0.044 (0.034) 0.006 (0.017) 0.862 (0.057) 0.721 (0.068) 0.244 (0.017) Yes Yes 14.12% 646923 3384 2.056 0.37% 0.52% Probability of Bankruptcy 0.612 (0.036) 0.638 (0.035) 1.213 (0.178) 1.125 (0.131) 0.035 (0.013) 1.823 (0.712) 1.900 (0.751) 0.181 (0.054) Yes Yes 18.66% 646923 632 6.686 0.18% 0.10%

ln(E) ln(F) 1/E rit1 rmt1 NI/TA CDS Firm CDS Active DD Time Fixed Eects Industry Fixed Eects Pseudo R2 N # of Downgrades(Bankruptcy) CDS Active Odds Ratio CDS Active Marginal Eect Sample Probability of a Downgrade(Bankruptcy)

57

Table A6 Impact of Credit Default Swaps Trading on Credit Quality: Distance-to-Default Matching
This table presents the estimates of the probability of credit downgrades/bankruptcy using a logistic model in a sample including rms with credit default swaps (CDS) and non-CDS distance-to-default (DD) matched rms. Each matched rm selected is the one rm with the closest DD to the target rm. DD is calculated from the Merton (1974) model. ln(E) is the logarithm of the rms equity value. ln(F) is the logarithm of the book value of the rms debt. 1/E is the inverse of the rms annualized equity volatility. rit1 rmt1 is the rms excess return over the past year, and NI/TA is the rms ratio of net income to total assets. To estimate the impact of CDS trading on the probability of credit downgrades/bankruptcy, we include CDS variables in the model specication. CDS Firm equals one if the rm is in the CDS sample and zero otherwise. CDS Active is a dummy variable that equals one if the rm has CDS traded on its debt one year before month t. The coecient of interest is that of CDS Active, which captures the impact of CDS trading on the probability of credit downgrades or bankruptcy after the inception of CDS trading. The sample period is 1997-2009, based on monthly observations. (*** Signicant at 1% level, ** signicant at 5% level, and * signicant at 10% level. The numbers in parentheses are standard errors.) Distance-to-Default Matching Probability of Downgrades Probability of Bankruptcy (1) (2) (3) (4) 0.462 0.447 0.923 0.891 (0.027) (0.028) (0.114) (0.113) 0.318 0.270 0.853 0.865 (0.030) (0.031) (0.116) (0.118) 0.155 0.008 1.905 1.971 (0.042) (0.038) (0.315) (0.317) 0.614 0.09 0.076 0.101 (0.073) (0.056) (0.191) (0.196) 0.845 0.700 0.331 0.994 (0.133) (0.221) (0.221) (0.259) 1.307 1.809 (0.100) (0.759) 0.586 1.313 2.196 0.773 (0.083) (0.069) (0.759) (0.299) Yes Yes Yes Yes Yes Yes Yes Yes 12.02% 8.03% 23.16% 23.05% 119143 119143 119143 119143 1469 1469 67 67 1.797 3.717 8.989 2.166 0.64% 1.46% 0.12% 0.04% 1.13% 1.14% 0.05% 0.05%

ln(E) ln(F) 1/E rit1 rmt1 NI/TA CDS Firm CDS Active Time Fixed Eects Industry Fixed Eects Pseudo R2 N # of Downgrades (Bankruptcy) CDS Active Odds Ratio CDS Active Marginal Eect Sample Probability of a Downgrade (Bankruptcy)

58

Table A7 Credit Rating and CDS Eects

This table investigates the impact of credit default swaps (CDS) trading on the probability of bankruptcy in subsamples of investment grade and non-investment grade rms. ln(E) is the logarithm of the rms equity value. ln(F) is the logarithm of the book value of the rms debt. 1/E is the inverse of the rms annualized equity volatility. rit1 rmt1 is the rms excess return over the past year, and NI/TA is the rms ratio of net income to total assets. To estimate the impact of CDS trading on the probability of credit downgrades/bankruptcy, we include CDS variables in the model specication. CDS Firm equals one if the rm is in the CDS sample and zero otherwise. CDS Active is a dummy variable that equals one if the rm has CDS traded on its debt one year before month t. The sample period is 1997-2009, based on monthly observations. (*** Signicant at 1% level, ** signicant at 5% level, and * signicant at 10% level. The numbers in the parentheses are standard errors.) Probability of Bankruptcy Investment Grade Non-investment Grade 0.705 0.704 (0.024) (0.024) 0.702 0.702 (0.023) (0.023) 1.825 1.625 (0.138) (0.134) 1.262 1.323 (0.110) (0.112) 0.036 0.037 (0.013) (0.013) 1.525 2.182 (1.004) (1.002) 1.893 2.721 (1.041) (1.024) Yes Yes Yes Yes 24.09% 23.64% 634895 608773 912 924 6.64 15.20 0.26% 0.40% 0.14% 0.15%

ln(E) ln(F) 1/E rit1 rmt1 NI/TA CDS Firm CDS Active Time Fixed Eects Industry Fixed Eects Pseudo R2 N # of Bankruptcies CDS Active Odds Ratio CDS Active Marginal Eect Sample Probability of Bankruptcy

Full Sample 0.713 (0.024) 0.711 (0.023) 1.626 (0.131) 1.320 (0.111) 0.038 (0.013) 2.009 (0.711) 2.373 (0.729) Yes Yes 24.18% 658966 940 10.73 0.33% 0.14%

59

Table A8 Mergers & Acquisitions and the CDS Eect

This table presents the estimates of the probability of bankruptcy using a logistic model in a sample excluding rms with a Mergers & Acquisitions (M&A) event. M&A data are obtained from SDC Interface. ln(E) is the logarithm of the rms market value of equity. ln(F) is the logarithm of the book value of the rms debt, where book debt is the sum of short-term debt and 50% of long-term debt. 1/E is the inverse of the rms annualized equity volatility. rit1 rmt1 is the rms excess return over the past year, and NI/TA is the rms ratio of net income to total assets. To estimate the impact of CDS trading on the probability of credit downgrades or bankruptcy, we include CDS variables in the model specication. CDS Firm equals one if the rm is in the CDS sample and zero otherwise. CDS Active is a dummy variable that equals one after and zero before the inception of CDS trading. The coecient of interest is that of CDS Active, which captures the impact of CDS trading on the probability of bankruptcy after the inception of CDS trading. The sample period is 1997-2009, based on monthly observations. (*** Signicant at 1% level, ** signicant at 5% level, and * signicant at 10% level. The numbers in parentheses are standard errors.) Probability of Bankruptcy 0.685 (0.025) 0.697 (0.024) 1.907 (0.148) 1.380 (0.121) 0.033 (0.014) 1.755 (0.712) 1.985 (0.735) Yes Yes 25.20% 563771 839 7.279 0.30% 0.15%

ln(E) ln(F) 1/E rit1 rmt1 NI/TA CDS Firm CDS Active Time Fixed Eects Industry Fixed Eects Pseudo R2 N # of Bankruptcy CDS Active Odds Ratio CDS Active Marginal Eect Sample Probability of a Bankruptcy

60

Table A9 Probability of Credit Default Swaps Trading: Additional Instruments

This table presents the estimates of the probability of credit default swaps (CDS) trading using a probit model. Propensity scores are estimated based on the model parameters. ln(Asset) is the logarithm of the rms total assets value. Leverage is dened as the ratio of book debt to the sum of book debt and market equity, where book debt is the sum of short-term debt and 50% of long-term debt and market equity is the measure of the number of common shares outstanding multiplied by the stock price. ROA is the rms return on assets. rit1 rmt1 is the rms excess return over the past year. Equity Volatility is the rms annualized equity volatility. PPENT/Total Asset is the ratio of property, plant and equipment to total assets. Sales/Total Asset is the ratio of sales to total assets. EBIT/Total Asset is the ratio of earnings before interest and tax to total assets. WCAP/Total Asset is the ratio of working capital to total assets. RE/Total Asset is the ratio of retained earnings to total assets. Cash/Total Asset is the ratio of cash to total assets. CAPX/Total Asset is the ratio of capital expenditure to total assets. Investment Grade is a dummy variable that equals one if the rm has an investment grade (BBB- or above) rating. Rated is a dummy variable that equals one if the rm is rated. Trace Coverage is a dummy that equals one for rms in the Trade Reporting and Compliance Engine (TRACE) database. Post CFMA is the a dummy that equals one for the period after the Commodity Futures Modernization Act of 2000 (CFMA). The Inverse Mills Ratio is calculated from the rst-stage probit regression modeling the probability of CDS trading. The sample period is 1997-2009, based on monthly observations. (*** Signicant at 1% level, ** signicant at 5% level, and * signicant at 10% level. The numbers in parentheses are standard errors.)

61

Trace Coverage Post CFMA Ln(Asset) Leverage ROA rit1 rmt1 Equity Volatility PPENT/Total Asset Sales/Total Asset EBIT/Total Asset WCAP/Total Asset RE/Total Asset Cash/Total Asset CAPX/Total Asset Investment Grade Rated Time Fixed Eects Industry Fixed Eects Pseudo R2 N #CDS Event

Panel A: Probability of CDS Trading Probability of CDS Trading CDS Prediction CDS Prediction Model 4 Model 5 0.512 (0.024) 0.386 (0.068) 0.799 0.797 (0.005) (0.005) 0.403 0.417 (0.025) (0.026) 0.020 0.012 (0.016) (0.016) 0.095 0.099 (0.010) (0.010) 0.055 0.068 (0.015) (0.015) 0.373 0.357 (0.029) (0.029) 0.022 0.021 (0.003) (0.003) 0.311 0.256 (0.060) (0.060) 0.144 0.159 (0.023) (0.024) 0.018 0.023 (0.005) (0.006) 0.249 0.251 (0.037) (0.037) 1.914 1.862 (0.114) (0.115) 0.944 0.916 (0.015) (0.013) 0.957 0.962 (0.015) (0.015) Yes Yes Yes Yes 38.79% 38.76% 690111 690111 551 551

62

Panel B: Treatment Eects Model with Instrumental Variables Probability of Bankruptcy CDS Prediction CDS Prediction Model 4 Model 5 ln(E) 0.639 0.639 (0.022) (0.022) ln(F) 0.646 0.645 (0.022) (0.022) 1/E 1.400 1.400 (0.125) (0.125) rit1 rmt1 1.330 1.330 (0.109) (0.109) NI/TA 0.032 0.032 (0.013) (0.013) CDS Firm 2.271 2.267 (0.710) (0.710) CDS Active 2.638 2.628 (0.747) (0.745) Inverse Mills Ratio 0.030 0.035 (0.128) (0.124) Time Fixed Eects Yes Yes Industry Fixed Eects Yes Yes Pseudo R2 22.42% 22.42% N 657438 657438 # of Bankruptcy 940 940 CDS Active Odds Ratio 13.985 13.846 CDS Active Marginal Eect 0.37% 0.37% Sample Probability of a Bankruptcy 0.14% 0.14%

63

Table A10 Impact of Credit Default Swaps Trading on Credit Quality: CDS Event
This table presents the estimates of the probability of bankruptcy using a logistic model. In contrast to the baseline results in Table II, we shift the CDS introduction date by one year as a falsication test. ln(E) is the logarithm of the rms equity value. ln(F) is the logarithm of the rms debt. 1/E is the inverse of the rms annualized equity volatility. rit1 rmt1 is the rms excess return over the past year, and NI/TA is the rms ratio of net income to total assets. The sample period is 1997-2009, based on monthly observations. (*** Signicant at 1% level, ** signicant at 5% level, and * signicant at 10% level. The numbers in the parentheses are standard errors.) Probability of Bankruptcy ln(E) 0.714 (0.024) ln(F) 0.712 (0.023) 1/E 1.627 (0.131) rit1 rmt1 1.321 (0.111) NI/TA 0.038 (0.013) CDS Firm 12.674 (168.47) CDS Active 12.959 (168.47) Time Fixed Eects Yes Industry Fixed Eects Yes Pseudo R2 24.20% N 658966 # of Bankruptcy 940 CDS Active Odds Ratio CDS Active Marginal Eect 1.79% Sample Probability of 0.14% Bankruptcy

64

Table A11 Impact of Credit Default Swaps Trading on Bankruptcy: The Feedback Mechanism
This table investigates the impact of credit default swaps (CDS) trading on a rms probability of bankruptcy, controlling for rms CDS spread status. Over Priced (Under Priced ) is a dummy that equals one for rms that are likely to be overpriced (underpriced), as measured by the basis between the CDS and bond yield spreads. ln(E) is the logarithm of the rms equity value. ln(F) is the logarithm of the rms debt. 1/E is the inverse of the rms annualized equity volatility. rit1 rmt1 is the rms excess return over the past year, and NI/TA is the rms ratio of net income to total assets. The sample period is 1997-2009, based on monthly observations. (*** Signicant at 1% level, ** signicant at 5% level, and * signicant at 10% level. The numbers in the parentheses are standard errors.) Probability of Bankruptcy Model 1 Model 2 Model 3 Model 4 ln(E) 0.696 0.696 0.696 0.697 (0.030) (0.030) (0.030) (0.030) ln(F) 0.715 0.714 0.714 0.715 (0.029) (0.029) (0.029) (0.029) 1/E 1.630 1.628 1.626 1.632 (0.188) (0.188) (0.188) (0.188) rit1 rmt1 1.750 1.747 1.750 1.749 (0.174) (0.174) (0.174) (0.174) NI/TA 0.042 0.042 0.042 0.042 (0.018) (0.018) (0.018) (0.018) CDS Firm 1.578 1.532 1.498 1.497 (1.005) (1.005) (1.005) (1.005) CDS Active 1.982 2.066 1.932 1.932 (1.021) (1.017) (1.021) (1.021) CDS Active*Over Priced 7.804 9.851 (211.54) (573.80) Over Priced 7.205 9.284 (211.54) (573.80) CDS Active*Under Priced 1.946 1.833 (613.85) (616.39) Under Priced 9.876 9.909 (486.90) (486.85) CDS Active*Mis-pricing 8.796 (195.73) Mis-pricing 8.410 (195.72) Time Fixed Eects Yes Yes Yes Yes Industry Fixed Eects Yes Yes Yes Yes Pseudo R2 27.54% 27.54% 27.56% 27.53% N 398638 398638 398638 398638 # of Bankruptcies 530 530 530 530 CDS Active Odds Ratio 7.257 7.893 6.903 6.903 CDS Active Marginal Eect 0.25% 0.26% 0.25% 0.25% Sample Probability of Bankruptcy 0.13% 0.13% 0.13% 0.13%

65

Table A12 Impact of Credit Default Swaps Trading on Bankruptcy: Analyst Coverage
This table investigates the impact of credit default swaps (CDS) trading on a rms probability of bankruptcy in a sample including rms with high (low) analyst coverage. Analyst coverage has been used as a proxy for the availability of private information. High (low) analyst coverage is indicated by the number of analysts of a rm being above (below) the median in the sample. ln(E) is the logarithm of the rms equity value. ln(F) is the logarithm of the rms debt. 1/E is the inverse of the rms annualized equity volatility. rit1 rmt1 is the rms excess return over the past year, and NI/TA is the rms ratio of net income to total assets. The sample period is 1997-2009, based on monthly observations. (*** Signicant at 1% level, ** signicant at 5% level, and * signicant at 10% level. The numbers in the parentheses are standard errors.) Probability of Bankruptcy Low Analyst Coverage High Analyst Coverage 0.596 0.713 (0.032) (0.024) 0.584 0.711 (0.032) (0.023) 1.773 1.626 (0.209) (0.131) 1.286 1.320 (0.156) (0.111) 0.026 0.038 (0.017) (0.013) 1.537 2.009 (1.006) (0.711) 1.986 2.373 (1.044) (0.729)

ln(E) ln(F) 1/E rit1 rmt1 NI/TA CDS Firm CDS Active CDS Active* Low Coverage Low Coverage Time Fixed Eects Industry Fixed Eects Pseudo R2 N # of Bankruptcies CDS Active Marginal Eect Sample Probability of Bankruptcy

Yes Yes 20.12% 256404 450 0.34% 0.18%

Yes Yes 28.71% 402562 490 0.32% 0.12%

Full Sample 0.712 (0.024) 0.710 (0.023) 1.660 (0.133) 1.319 (0.111) 0.039 (0.013) 2.021 (0.711) 2.329 (0.737) 0.134 (0.359) 0.129 (0.070) Yes Yes 24.21% 658966 940 0.32% 0.14%

66